I have a hard time believing their numbers. If you can pay off a mac mini in 2-4 months, and make $1-2k profit every month after that, why wouldn’t their business model just be buying mac minis?
The numbers are optimistically legit -- it's calculated based purely considering we have demand for all machines at all times. We don't have that right now, but fairly optimistic that people will do it.
That's why we don't recommend purchasing a new machine. Existing machine is no cost for you to run this.
Electricity is one cost, but it will get paid off from every request it receives. Electricity is only deducted when you run an inference. If you have any questions, DM me @gajesh on Twitter.
> That's why we don't recommend purchasing a new machine. Existing machine is no cost for you to run this.
You misunderstood. If the ROI is there, there is enough capital in existence for you to accelerate your profit. So why even deal with the complexity of renting people's hardware when you can do it yourself?
You're not taking into account the thermal strain on the machine, though. A machine that's 100% utilized (even worse if it's in bursts) will last less than an idle machine.
Not appreciably, and not before a 5-yr AppleCare+ warranty expires.
Out of our >3000 currently active Apple Silicon Macs, failures due to non-physical damage are in the single digits per year. Of those, none have been from production systems with 24/7 uptime and continuous high load, which reflects your parenthetical.
Perhaps we haven't met the other end of the bathtub curve yet, but we also won't be retaining any of these very far beyond their warranty period, much less the end of their support life.
AppleCare+ annual is perpetual as long as you keep paying it (and Apple offers to switch to that when your 3-year lump sum expires if you choose that instead). I’m guessing it ends whenever they officially discontinue hardware support, which has traditionally been about 7 years after the last unit is produced, but I haven’t reached that yet to know for sure.
I think the point of this is more "use the machine you have at home" than "do a TCO analysis and see if it's profitable", though. People like to keep their machines working for longer, generally.
> A machine that's 100% utilized (even worse if it's in bursts) will last less than an idle machine.
How much though? Say I have three Mac Minis next to each other, one that is completely idle but on, one that bursts 100% CPU every 10 minutes and one that uses 100% CPU all the time, what's the difference on how long the machines survives? Months, years or decades?
> Existing machine is no cost for you to run this.
That is not at all how modern chips work. Idle chips are mostly powered down, non-idle ones are working and that causes real measurable wear and tear on the silicon. CPU, RAM, NAND all wear and tear measurably with use on current manufacturing processes.
The question is, do they wear faster than they become obsolete, as in much more expensive to run than buying a new one with higher compute/watt. (and you can also factor in the ability to run latest models at usable speed)
It’s complicated. When you design with modern PDKs, you consider the expected duty cycle of the device, expected temps, and the wear and tear on the silicon all together. That affects the layout of the chip as well as certain choices about widths of various features. Generally, one designs consumer SoCs to last 10 years with the expected duty cycle (low). With more wear you could run out of your “years" much faster, maybe even before the warranty.
I don't worry about bandwidth or constant CPU use, but the one thing that will kill my mac is burning out the SSD.
The calculator gives numbers for nearly everything, but I can't obviously see how much space it needs for model storage or how many writes of temp files I should expect if I'm running flat out.
put that stuff on an external disk perhaps, it will eventually crater, but it's easier to replace than macbook internal storage (how are they doing mac minis these days?)
If you start buying minis, then you need to house, power, and cool them. So you are building a mini data center. If you are building a small data center, economies of scale will drive you to want to build larger and larger. However, this gets expensive and neighbors tend to not like data centers (for good reason). To me this seems like asymmetric warfare against hyper-scalers.
Because they don’t have that much initial money in their pocket, while the idle computer is already there, and the biggest friction point is convincing people to install some software. Both producing rhetoric and software are several order of magnitude cheaper than to directly own and maintain a large fleet of hardware with high guarantee of getting the electrical stable input in a safe place to store them.
Assuming that getting large chunk of initial investment is just a formality is out of touch with 99% of people reality out there, when it’s actually the biggest friction point in any socio-economical endeavour.
No provider maintains 100% utilization of GPUs at full rate. Demand is bursty - even if this project is successful, you might expect, e.g., things to be busy during the stock market times when Claude is throwing API errors and then severely underutilized during the same times that Anthropic was offering two-for-one off peak use.
And then there's a hit for overprovisioning in general. If the network is not overprovisioned somewhat, customers won't be able to get requests handled when they want, and they'll flee. But the more overprovisioned it is, the worse it is for compute seller earnings.
I suspect an optimistic view of earnings from a platform like this would be something like 1/8 utilization on a model like Gemma 4. Their calculator estimates my m4 pro mini could earn about $24/month at 3 hours/day on that model. That seems plausible.
Solid q. I think the part of it is that it’s really easy to attract some “mass” (capital) of users, as there are definitely quite a few of idle Macs in the world.
Non-VC play (not required until you can raise on your own terms!) and clear differentiation.
If you want to go full-business-evaluation, I would be more worried about someone else implementing same thing with more commission (imo 95% and first to market is good enough).
I think the point they’re making though is that the numbers seem too good to be true.
ie. Does anyone know the payback time for a B100 used just for inference? I assume it’s more than a couple of months? Or is it just training that costs so much?
Or like anything else it will be too good to be true at the very beginning but then once people hear about it and it gets popular supply overtakes demand and the mac minis go back to being idle most of the day.
When YouTubers start making videos about it you know it's too late.
The question is what is the hassle of running it plus wear and tear. Max price will tend to that. It is not like crypto where there is capital investment in a rig that can do nothing else. People are using their existing laptop. So I reckon 20-50 a year max per laptop.
The numbers are obviously high, because if this takes off then the price for inference will also drop. But I still think it’s a solid economic model that benefits low income countries the most. In Ukraine, for example, I know people who live on $200/month. A couple Mac Minis could feed a family in many places.
As a business owner, I can think of multiple reasons why a decentralized network is better for me as a business than relying on a hyperscaler inference provider. 1. No dependency on a BigTech provider who can cut me off or change prices at any time. I’m willing to pay a premium for that. 2. I get a residential IP proxy network built-in. AI scrapers pay big money for that. 3. No censorship. 4. Lower latency if inference nodes are located close to me.
They already have an iPhone. They could save up or borrow for a Mac Mini if they had to. Some of those people I know who live on $200/month have $30k in the bank.
Not really. There are lots of people who have low income and low spending, but not low savings. Retired pensioners with savings. Young families who inherited from deceased parents/grandparents. Highly paid professionals on sabbatical. I've met people from all of those categories in Ukraine who live on $200/month.
Isn't this same premise as "lets buy few GPUs to mine crypto and have passive income"?
It didn't work very well and it probably won't work now either. If there is money to be made, bigger players will get in there, buy out all mac minis they can, drive price up for regular people and inevitably drive inference price down so you'll be lucky to get initial investment back
No it's not the same premise at all. Crypto doesn't do anything useful for legitimate businesses. AI inference is very useful for legitimate businesses, and so are residential IP proxies for scraping. And by definition, residential IPs cannot be centralized. And as building GPUs becomes more expensive, the existing pool of second hand unused hardware becomes more valuable, not less. The problem with crypto mining is that it quickly becomes unprofitable for small scale deployments. I'm not sure if AI inference would be, especially for the decentralized benefits of lower latency.
It is the same premise, because the person you are responding to is not talking about the moral implications at all, only about the financial / hardware implications.
Running AI inference increases the power draw, and requires certain hardware.
Mining bitcoin increases the power draw, and requires certain hardware.
OP's point thus stands: Bad players will find places to get far cheaper power than the intended audience, and will buy dedicated hardware, at which point the money you can earn to do this will soon drop below the costs for power (for folks like you and me).
Maybe that won't happen, but why won't that happen?
The main problem with crypto is there is no universal need for it. The demand for crypto doesn’t keep increasing as compute gets cheaper. But the demand for AI inference is only growing, and making it cheaper would likely only increase demand. So it’s not a race to the bottom. Sure hyper focused players can earn more at higher margins. But average players can probably still earn decently. Take for example electricity. It can still be profitable for a home in Germany to install balcony solar and make a little money selling back to the grid even though it’s obviously not as efficient as an industrial power plant. Mom and pop AI inference don’t have to be super efficient as long as they serve a universal need - it will be like balcony solar in Europe.
On the latency point - your requests are still going through the coordinator of the system here. So on average strictly worse than a large provider.
You - Darkbloom - Operator - Darkbloom - you, vs
You - Provider - you
---
On the censorship point - this is an interesting risk surface for operators. If people are drawn my decentralized model provisioning for its lax censorship, I'm pretty sure they're using it to generate things that I don't want to be liable for.
If anything, I could imagine dumber and stricter brand-safety style censorship on operator machines.
I'm not talking about Darkbloom specifically, but rather this business model in general. I'm sure a future version of Darkbloom could be P2P for better latency. Or their central operator nodes could be geo-balanced. Liability for censorship doesn't matter if it's truly zero trust. Anyway censorship is not my main concern. Low-latency decentralized inference with no US BigTech dependency is a much bigger selling point in Europe.
> These are estimates only. We do not guarantee any specific utilization or earnings. Actual earnings depend on network demand, model popularity, your provider reputation score, and how many other providers are serving the same model.
Of course these numbers are ridiculous. Mac Mini (let's assume Apple releases M5 Pro) tops Int8 (let's assume it is the same as FP8, which it is not) at ~50 TFLOPs, with Draw Things, we recently developed hybrid NAX + ANE inference, which can get you ~70 TFLOPs.
A H200 gives you ~4 PFLOPs, which is ~60x at only ~40x price (assuming you can get a Mac Mini at $1000). (Not to mention, BTW, RTX PRO 6000 is ~7x price for ~40x more FLOPs).
There are many people who do not have ready access to a million dollars to purchase said Mac minis, much less the operating capital to rack & operate them.
Very smart play to build a platform, get scale, and prove out the software. Then either add a small network fee (this could be on money movement on/off platform), add a higher tier of service for money, and/or just use the proof points to go get access to capital and become an operator in your own pool.
If those numbers are true, they could tart with one Mac and can double every few months. But, I guess there are also many people who do not have ready access to whatever a Mac mini costs either...
Because the "ship software to people, rent their hardware" model has zero up front investment required, presumably. And they don't have to deal with power, cooling, real estate.
I guess if it only works at scale capital is maybe the answer. Like enough cash to buy 5 or 10 or even 100 minis seem doable - but if the idea only works well when you have 10,000 running - that makes some sense.
Because their numbers don’t work out. When you do the math on token cost versus inference speed, you get something that barely breaks even even with cheap power.
Also they’ve already launched a crypto token, which is a terrible sign.
I installed this so you don't have to. It did feel a bit quirky and not super polished. Fails to download the image model. The audio/tts model fails to load.
In 15 minutes of serving Gemma, I got precisely zero actual inference requests, and a bunch of health checks and two attestations.
At the moment they don't have enough sustained demand to justify the earning estimates.
I kind of see your point, but I also kind of don't.
Sure, it would be great if you'd immediately get hammered with hundreds of requests and start make money quickly. It would also be great if it was a bit more transparent, and you could see more stats (what counts as "idle"? Is my machine currently eligible to serve models?). But it's still very new, I'd say give it some time and let's see how it goes.
If you have it running and you get zero requests, it uses close to zero power above what your computer uses anyway. It doesn't cost you anything to have it running, and if you get requests, you make money. Seems like an easy decision to me.
Bootstrapping will be near-impossible (or incredibly costly) unless they offer inference consumers models with established demand arriving at some least-cost router service where they can undercut the competition (if they actually can). And then dogfood the opportunistic provider side on their own Macs, but with a preference to putting third parties first in the queue. Everything else is just wishful thinking.
weird to learn that they do not generate inference requests to their network themselves to motivate early adopters at least to host their inference software
If they paid promised > $1k/m for FLUX 2B on a Mac they would go broke in less than a month. On a single 5090 that model would provide an inference througput so high they'd have to pay close to $50k/m for the results.
The numbers are absolute fraud. You shouldn't be installing their software cause fraud could be not just about numbers.
Can you rephrase that? I don't think I've read it correctly. It sounds like you are saying it would normally cost $50k on a 5090 and they can do equivalent work paying $1k. That's sounds like a $49k profit margin, but you say they will go broke.
Given their estimates of a Mac being able to generate $1k (per month?) a 5090 with a lot more power would be able to generate $50k. For a $3k piece of hardware. Which is obviously not realistic. (As in, nobody is paying that much for the images, which seems to match well with no actual requests on the system.)
and I don't think they ever will unless they're highly competitive (hopefully that price they have stays? at least for users)
I was thinking of building this exact thing a year ago but my main stopper was economics: it would never make sense for someone to use the API, thus nobody can make money off of zero demand.
I guess we just have to look at how Uber and Airbnb bootstrapped themselves. Another issue with my original idea was that it was for compute in general, when the main, best use-case, is long(er)-running software like AI training (but I guess inference is long running enough).
But there already exist software out there that lets you rent out your GPU so...
People underestimate how efficient cost/token is for beefy GPUs if you are able to batch. Unlikely for one off consumer unit to be able to compete long term.
You can see in their stats view they have a lot of providers/nodes connected but practically no actual demand/consumers. They just launched and I'm sure get providers was top of their agenda, but it's essentially unusable as a provider unless they perform some serious lift to get actual paying customers.
I received the same error, but it was followed by this line in the logs, which might explain the lack of inference requests assume there is actual demand...
WARN STT backend failed health check — model will NOT be advertised
I disabled the audio model (had to remove it, it was so buggy) and then started up with a text only model. Serving started without error. The system simply has no requests. The economics seem like a mirage anyway.
You have to install their MDM device management software on your computer. Basically that computer is theirs now. So don't plan on just handing over your laptop temporarily unless you don't mind some company completely owning your box. Still might be a validate use for people with slightly old laptops lying around, but beware trying to share this computer with your daily activities if you e.g. use a bank on a browser on this computer regularly. MDM means they can swap out your SSL certs level of computer access, please correct me if I'm wrong.
MDMs on macOS are permissioned via AccessRights, and you can verify that their permission set is fairly minimal and does not allow what you've described here (bits 0, 4, 10).
That said, their privacy posture at the cornerstone of their claims is snake oil and has gaping holes in it, so I still wouldn't trust it, but it's worth being accurate about how exactly they're messing up.
You are right - the "nonce binding" the paper uses doesn't seem convincing. The missing link is that Apple's attestation doesn't bind app generated keys to a designated requirement, which would be required to create a full remote attestation.
> If you can prove a public key is generated by the SEP of a machine running with all Apple's security systems enabled, then you can trivially extend that to confidential computing because the macOS security architecture allows apps to block external inspection even by the root user.
It only effectively allows this for applications that are in the set of things covered by SIP, but not for any third-party application. There's nothing that will allow you to attest that arbitrary third-party code is running some specific version without being tampered with, you can only attest that the base OS/kernel have not been tampered with. In their specific case, they attempt to patch over that by taking the hash of the binary, but you can simply patch it before it starts.
To do this properly requires a TEE to be available to third-party code for attestation. That's not a thing on macOS today.
I wiped my post because you are right. I don't think it needs a full SGX-style TEE. What's missing is a link to designated requirements. Abusing a nonce field doesn't seem to work, or if it does I can't figure out how. The MDM/MDA infrastructure would need to be able to include:
public key from SEP -> designated requirement of owning app binary
The macOS KeyStore infrastructure does track this which is why I thought it'd work. But the paper doesn't mention being able to get this data server side anywhere. Instead there's this nonce hack.
It's odd that the paper considers so many angles including things like RDMA over Thunderbolt, but not the binding between platform key and app key.
Reading the paper again carefully I get the feeling the author knows or believes something that isn't fully elaborated in the text. He recognizes that this linkage problem exists, proposes a solution and offers a security argument for it. I just can't understand the argument. It appears APNS plays a role (apple push notification service) and maybe this is where app binding happens but the author seems to assume a fluency in Apple infrastructure that I currently lack.
I can buy the idea that if you can have the MDM infrastructure attest the code signing identity through the designated requirements, that you can probably come pretty close, but I'm still not quite sure you get there with root on macOS (and I suspect that this is part of why DCAppAttest hasn't made it to macOS yet).
Certainly, it still doesn't get you there with their current implementation, as the attempts at blocking the debugger like PT_DENY_ATTACH are runtime syscalls, so you've got a race window where you can attach still. Maybe it gets you there with hardened runtime? I'd have to think a bit harder on that.
Yeah I didn't quite understand the need for PT_DENY_ATTACH. Hardened runtime apps that don't include get-task-allow are already protected from debugger attach from the start of the process, unless I misunderstood something.
I'm not quite sure why Apple haven't enabled DCAppAttest on macOS. From my understanding of the architecture, they have every piece needed. It's possible that they just don't trust the Mac platform enough to sign off on assertions about it, because it's a lot more open so it's harder to defend. And perhaps they feel the reputational risk isn't worth it, as people would generalize from a break of App Attest on macOS to App Attest on iOS where the money is. Hard to say.
They can never have supervised permissions on your device unless your mac was wiped and registered in Apple Business Manager. As a BYOD, it cannot be locked.
Whether it makes sense to enroll a mac as a BYOD is another question.
Unfortunately, verifiable privacy is not physically possible on MacBooks of today. Don't let a nice presentation fool you.
Apple Silicon has a Secure Enclave, but not a public SGX/TDX/SEV-style enclave for arbitrary code, so these claims are about OS hardening, not verifiable confidential execution.
It would be nice if it were possible. There's a lot of cool innovations possible beyond privacy.
I wrote a whole SDK for using SGX, it's cool tech. But in theory on Apple platforms you can get a long way without it. iOS already offers this capability and it works OK.
macOS has a strong enough security architecture that something like Darkbloom would have at least some credibility if there was a way to remotely attest a Mac's boot sequence and TCC configuration combined with key-to-DR binding. The OS sandbox can keep apps properly separated if the kernel is correct and unhacked. And Apple's systems are full of mitigations and roadblocks to simple exploitation. Would it be as good as a consumer SGX enclave? Not architecturally, but the usability is higher.
As if you get privacy with the inference providers available today? I have more trust in a randomly selected machine on a decentralized network not being compromised than in a centralized provider like OpenAI pinky promising not to read your chats.
I would say the chances of OpenAI itself getting hacked and your secrets in logs getting leaked are about the same or less as the chances of a randomly selected machine on a decentralized network being reverse-engineered by a determined hacker. There's no risk-free option, every provider comes with risks. If you care about infosec you have to do frequent secret rotation anyway.
They use the TEE to check that the model and code is untampered with. That's a good, valid approach and should work (I've done similar things on AWS with their TEE)
The key question here is how they avoid the outside computer being able to view the memory of the internal process:
> An in-process inference design that embeds the in-
ference engine directly in a hardened process, elimi-
nating all inter-process communication channels that
could be observed, with optional hypervisor mem-
ory isolation that extends protection from software-
enforced to hardware-enforced via ARM Stage 2 page
tables at zero performance cost.[1]
I was under the impression this wasn't possible if you are using the GPU. I could be misled on this though.
This entire paper smells of LLM, I'm sure even the most distinguished academic would refrain from using notation to prove that the SIP status cannot change during operation.
While they do make this argument, realistically anyone sending their prompt/data to an external server should assume there will be some level of retention.
And more so in particular, anyone using Darkbloom with commercial intents should only really send non-sensitive data (no tokens, customer data, ...) I'd say only classification tasks, imagine generation, etc.
> PT_DENY_ATTACH (ptrace constant 31): Invoked
at process startup before any sensitive data is loaded.
Instructs the macOS kernel to permanently deny all
ptracerequests against this process, including from
root. This blocks lldb, dtrace, and Instruments.
> Hardened Runtime: The binary is code-signed with
hardened runtime options and explicitly without the
com.apple.security.get-task-allow
entitlement. The kernel denies task_for_pid()
and mach_vm_read()from any external process.
> System Integrity Protection (SIP): Enforces both of
the above at the kernel level. With SIP enabled, root
cannot circumvent Hardened Runtime protections, load
unsigned kernel extensions, or modify protected sys-
tem binaries. Section 5.1 proves that SIP, once verified,
is immutable for the process lifetime.
Looking at their paper at [1], there's a gaping hole: there's no actual way to verify the contents of the running binaries. The binary hash they include in their signatures is self-reported, and can be modified. That's simply game over.
A note, as others have posted on this thread: I mention this as a concrete and trivial flaw in their whole strategy, but the issue is fundamental: there's no hardware enclave for third-party code available to do the type of attestation that would be necessary. Any software approach they develop will ultimately fall to that hole.
Apple is perfectly capable of doing remote attestation properly. iOS has DCAppAttest which does everything needed. Unfortunately, it's never been brought to macOS, as far as I know. Maybe this MDM hack is a back door to get RA capabilities, if so it'd certainly be intriguing, but if not as far as I know there's no way to get a Mac to cough up a cryptographic assertion that it's running a genuine macOS kernel/boot firmware/disk image/kernel args, etc.
It's a pity because there's a lot of unique and interesting apps that'd become possible if Apple did this. Darkbloom is just one example of what's possible. It'd be a huge boon to decentralization efforts if Apple activated this, and all the pipework is laid already so it's really a pity they don't go the extra mile here.
> If you read isSupported from an app running on a Mac device, the value is false. This includes Mac Catalyst apps, and iOS or iPadOS apps running on Apple silicon.
You can probably just tap the HTTP(S) connection to spy on the data coming through. I think it's a mistake to assume any kind of privacy for this service.
The biggest argument for remote attestation I can think of is to make sure nobody is returning random bullshit and cashing in prompt money on a massive scale.
Well. Running your machine to do inference will utilize more than 50W sustained load, I'd say more than double that. Plus electricity is more expensive here (but granted, I do have solar panels). Plus don't forget to factor in that your hardware will age faster.
Your hardware will age slower if you have consistent load.
Thermal stress from bursty workloads is much more of a wearing problem than electromigration. If you can consistently keep the SoC at a specific temperature, it'll last much longer.
This is also why it was very ironic that crypto miner GPUs would get sold at massive discounts. Everyone assumed that they had been ran ragged, but a proper miner would have undervolted the card and ran it at consistent utilization, meaning the card would be in better condition than a secondhand gamer GPU that would have constantly been shifting between 1% to 80% utilization, or rather, 30°C to 75°C
Their estimate is based on significantly lower consumption when under load. E.g. 25W for an M4 Pro mac mini. I have no idea if that’s realistic - but the m4s are supposedly pretty efficient (https://www.jeffgeerling.com/blog/2024/m4-mac-minis-efficien...)
Their example big earner models are FLUX.2 Klein 4B and FLUX.2 Klein 9B, which i imagine could generate a lot more tokens/s than a 26B model on your machine.
I have no idea if that is a good estimate of how much an M5 Pro can generate - but that’s what it says on their site.
They do a bit of a sneaky thing with power calculation: they subtract 12Ws of idle power, because they are assuming your machine is idling 24/7, so the only cost is the extra 18W they estimate you’ll use doing inference. Idk about you, but i do turn my machine off when i am not using it.
OpenAI has only about 5% paying customers, how does it generate revenue?
I don’t think this is a sustainable business model. For example, Cubbit tried to build decentralised storage, but I backed out because better alternatives now exist, and hardware continues to improve and become cheaper over time.
Your electricity and ownership are going to get lower return and does not actually requce CO2.
Yeah, only way to get there is assuming they're not giving prompt caching discounts while my laptop is getting prompt caching benefits, with very many large prompts. So yes I am skeptical of their numbers.
Any idea what makes for such a diff between your and theirs numbers? Batching? Or could they do a crazy prefix caching across all nodes to reduce the actual processing.
You're printing everyone's serial numbers publicly. https://console.darkbloom.dev/providers then "Security Verification" for any machine and then "Verify this device independently" -- all of this can be scraped.
This is one of those ideas I think makes perfect sense, but requires so much operational change for the entire stack, that it would be very difficult to scale:
- Convincing labs to run distributed, burst-y inference
- Convincing people to run their Mac all day, hoping to make a little profit
- Convincing users to trust a distributed network of un-trusted devices
I had a similar idea, pre-AI, just for compute in general. But solving even 1 of those 3 (swap AI lab for managed-compute-type-company, eg Supabase, Vercel) is nearly impossible.
...or convincing operators that jobs sent to their machines are legal, legitimate, and non-nefarious.
I could not find disclosure on their site about the guard-railing or safety-systems at the point the prompt is gathered from users which would intercept, log & prevent bad actors from inadvertently involving me in something illegal or immoral as an operator. Perhaps that disclosure exists and I just need to be linked to it; that would be welcome.
Having strong SETI@Home vibes from 25 years ago, except of course, this is not for the greater good of humanity, but a for-profit project.
Problem is, from a technical point of view, what kind of made sense back then (most people running desktops, fans always on, energy saving minimal) is kind of stupid today (even if your laptop has no fan, would you want it to be always generating heat?)...
I definitely want my laptops to be cool, quiet and idle most of the time.
I some times play about with local models via ollama/comfyui and more recently ace-step to generate music.
This is short bursts of heat 5-10 m during the render I would not be happy with that for multiple hours a day. I am sure that would have a negative effect on battery health.
@eigengajesh - Your cost estimator lists Mac Mini M4 Pro with only 24 or 48GB options, but the M4 Pro mini can also be configured with 64GB. At least, I hope so, as I'm typing this on one. ;-)
Oh, also, you seem to have some bugs:
Gemma:
WARN [vllm_mlx] RuntimeError: Failed to load the default metallib. This library is using language version 4.0 which is not supported on this OS. library not found library not found library not found
cohere:
2026-04-16T14:25:10.541562Z WARN [stt] File "/Users/dga/.darkbloom/bin/stt_server.py", line 332, in load_model
2026-04-16T14:25:10.541614Z WARN [stt] from mlx_audio.stt.models.cohere_asr import audio as audio_mod
2026-04-16T14:25:10.541643Z WARN [stt] ModuleNotFoundError: No module named 'mlx_audio.stt.models.cohere_asr'
Trying to download the flux image models fails with:
curl: (56) The requested URL returned error: 404
darkbloom earnings does not work
your documentation is inconstent between saying 100% of revenue to providers vs 95%
I think .. this needs a little more care and feeding before you open it up widely. :) And maybe lay off the LLM generated text before it gets you in trouble for promising things you're not delivering.
Had it for 3 days running with nary a request. Perhaps it's the chosen model that I am serving (Gemma 4 26b)? I did see some WARN logs right after startup but the description don't suggest any problems that would block accepting requests or processing them.
I'd love a way to do this locally -- pool all the PCs in our own office for in-office pools of compute. Any suggestions from anyone? We currently run ollama but manually manage the pools
If you set CPUSchedulingPolicy=idle Nice=19 IOSchedulingClass=idle in the ollama server configuration it should run in the background with lowest priority.
Seems like so much more work than "just" paying for https://huggingface.co or whichever other neocloud who already did all the setup for you and just waits for your credit card per minute/seconds/token.
Doubt this kind of workloads would agree to send data then to a cloud of randos devices, precisely when cloud providers to certify they aren't looking at clients data (Customer-managed encryption keys, CMEK).
Interesting concept. Two sided marketplaces are hard to bootstrap but maybe just enough curiosity would get the flywheel going. Hell they should just try and convince people to enroll as providers but then also use the service even if it’s hitting their own machines until there’s some degree of supply and demand pressure then try and get only providers to sign up. Or set up some way to encourage providers to promote others to use the service (the 100% rev share kind of breaks that concept but anything can change).
I wish this was self hostable, even for a license fee. Many businesses have fleets of Macs, sometimes even in stock as returned equipment from employees. Would allow for a distributed internal inference network, which has appeal for many orgs who value or require privacy.
I'm not sure how the economics works out. Pricing for AI inference is based on supply/demand/scarcity. If your hardware is scarce, that means low supply; combine with high demand, it's now valuable. But what happens if you enable every spare Mac on the planet to join the game? Now your supply is high, which means now it's less valuable. So if this becomes really popular, you don't make much money. But if it doesn't become somewhat popular, you don't get any requests, and don't make money. The only way they could ensure a good return would be to first make it popular, then artificially lower the number of hosts.
Love the concept, with some similarity to folding@home, though more personal gain.
But trying it out it still needs work, I couldn't download a model successfully (and their list of nodes at https://console.darkbloom.dev/providers suggests this is typical).
And as a cursory user, it took me some digging to find out that to cash out you need a Solana address (providers > earnings).
"Debugger attachment is blocked. Memory inspection is blocked." - reminds me old crackme challenges. Everything they mention can be bypassed, so determined person can start stealing data from the network. For me this is a killer of such distributed compute ideas, but who knows, maybe the non-enteprise users will be desperate enough for cheap compute to make this idea valid.
The hardware-attested privacy path is the interesting part of this, but the economic side has a quieter risk the thread has not named: the load tax per request.
MiniMax M2.5 239B from your catalog still has to load all 239B weights even though only 11B are active — that is roughly 120GB at Q4_K_M, and cold load from SSD on Apple Silicon is measurable in tens of seconds. Even the Qwen3.5 122B MoE lands around 65GB cold. If the coordinator routes request number two to a different idle Mac than request number one, or if the owner's machine spun the model out to free memory in between, each request pays that cold load before the first token. Keeping the model resident 24/7 solves the latency but eats into the power budget the operator is trying to amortize in the first place.
How does the coordinator decide which provider to keep warm for which model? A 16GB or 32GB home Mac cannot host Qwen3.5 122B MoE at all, and the Mac Studios that can are a much smaller slice of the 100M machine estimate.
Interesting to see an offering with this heritage [1] proposing flat earnings rates for inference operators here, rather than trying to sell a dynamic marketplace where operators compete on price in real-time.
Right now the dashboards show 78 providers online, but someone in-thread here said that they spun one up and got no requests. Surely someone would be willing to beat the posted rate and swallow up the demand?
I expect this is a migration target, but a tactical omission from V1 comms both for legitimate legibility reasons (I can sell x for y is easier to parse than 'I can participate in a marketplace') and slightly illegitimate legibility reasons (obscuring likely future price collapse).
Still - neat project that I hope does well.
[1] Layer Labs, formerly EigenLayer, is company built around a protocol to abstract and recycle economic security guarantees from Ethereum proof of stake.
Three binaries and a Python file:
darkbloom (Rust)
eigeninference-enclave (Swift)
ffmpeg (from Homebrew, lol)
stt_server.py (a simple FastAPI speech-to-text server using mlx_audio).
The good parts:
All three binaries are signed with a valid Apple Developer ID and have Hardened runtime enabled.
Bad parts:
Binaries aren't notarized. Enrolls the device for remote MDM using micromdm. Downloads and installs a complete Python runtime from Cloudflare R2 (Supply chain risk). PT_DENY_ATTACH to make debugging harder. Collects device serial numbers.
> Apple’s attestation servers will
only generate the FreshnessCode for a genuine device that
checks in via APNs. A software-only adversary cannot
forge the MDA certificate chain (Assumption 3). Com-
bined with SIP enforcement (preventing binary replace-
ment) and Secure Boot (preventing bootloader tampering),
this provides strong evidence that the signing key resides
in genuine Apple hardware.
My understanding from the paper is that doing so should cause certain things in Apple's hardware security enclaves to break a signing chain, and a server-side MDM system integrated with Apple servers can detect this. But I'm not familiar with the underlying technology, so not sure if underlying assumptions are incorrect.
Epyc has that VM encrypted memory thing, which comes pretty close. It does raise an interesting question, though: would a PCIe card passed through to a VM be able to DMA access the memory of neighboring devices?
"These are estimates only. We do not guarantee any specific utilization or earnings. Actual earnings depend on network demand, model popularity, your provider reputation score, and how many other providers are serving the same model.
When your Mac is idle (no inference requests), it consumes minimal power — you don't lose significant money waiting for requests. The electricity costs shown only apply during active inference.
Text models typically see the highest and most consistent demand. Image generation and transcription requests are bursty — high volume during peaks, quiet otherwise."
It's a good project that makes sense. I recommend adding a contractual layer as well, since it's free and makes sense. Operators could legally sign that they will not look into the inference layer. After all, the operators already have a financial relationship with this provider, so it makes sense to add a contract to it and keep operators from looking into other people's data that way, too. I wish this project a lot of success.
Wasn't there an idea about 15 years ago where you would open your browser, go to a webpage and that page would have a JavaScript based client that would run distributed workloads?
I believe the idea was that people could submit big workloads, the server would slice them up and then have the clients download and run a small slice. You as the computer owner would then get some payout.
I used to work at Distributive (formerly "Kings Distributed Systems") on its DCP compute platform" which is entirely what you're describing. You can deploy a JS/WASM based workload, and it will be "sliced" and served to browser-based compute nodes. With WebGPU you can sort of have inference executing in the browser too. Incredible people there with an awesome project
I added Python execution support via Pyodide (cpython compiled to wasm) and worked on a bunch of other random stuff like WebLLM inferencing during my time there.
Apart from Distributive, there's also the "Golem network", "Salad", "Koii" and various other similar projects.
---
I'm not sure if I'm convinced by the "Uber for compute" use case with compute buyers and compute workers (sellers), but if you're a university and you have 1000 Windows machines across all your computer labs, it'd be nice to leverage that compute for running research or something idk - especially with the price of ram / cloud offerings these days...
> but if you're a university and you have 1000 Windows machines across all your computer labs, it'd be nice to leverage that compute for running research or something idk - especially with the price of ram / cloud offerings these days...
This reminds me of the DevOps guy who made the developer laptops part of a Jenkins "swarm" under the thought that the machines were beefy and underutilized most of the time.
I think it’s important that systems like this exist, but getting them off the ground is non-trivial.
We’ve been building something similar for image/video models for the past few months, and it’s made me think distribution might be the real bottleneck.
It’s proving difficult to get enough early usage to reach the point where the system becomes more interesting on its own.
Curious how others have approached that bootstrap problem. Thanks in advance.
As one of the only people running a Mac Studio M3 Ultra with 512 GB of RAM on the network, I can tell you at sustained 100% GPU utilization I am measuring 250 watts max (at the power outlet). My solar panels are easily producing this. The power calculation goes away once you connect a solar panel. You can get a 400 watt solar panel on Amazon for $300.
> You can get a 400 watt solar panel on Amazon for $300.
Too expensive. It's probably producing 200 watts average for 8 hours a day. That's 1600 watt hours, which is about $1.60 at PG&E prices. That would take 187 days to recoup the cost of just the panel.
If you include installation costs and "what PG&E steals if you wire it to the same grid" it's probably more like 4x that, which is too long.
Tell me when we can have 400 watt solar panels for $50. Stupid capitalism literally forces solar panel prices to make it unprofitable.
People should never have to take out loans for solar. Solar should be subsidized and forced by the government to be so cheap that it repays for its cost within a month. Then we're talking. Most things I buy to save money, I expect them to repay within a month. Maybe 2 months max.
I'm not "worried" about that cost, I would rather just pay PG&E electricity if the solar panel cost $300.
Just pointing out why capitalism + solar is a failure. Capitalism reprices the good thing to be equally expensive to the bad thing, so that nobody buys the good thing anymore.
client side of this kind of needs to be open source unless I'm running it on a dedicated machine and firewalling it from the rest of my network. Or the company needs to have a very strong reputation and certifications. curlbash and go is a pretty hard sell for me
Cool idea, though hats off to anyone who got cohere-transcribe to show up as serving the model. I could get device to show up, but kept having issues getting their server to properly serve the model though it could just be the device I tested.
How can one do this safely? If I create a new, non-sudo user, can I install the MDM profile only for that user? I don't understand how this all works obviously so maybe this is a very dumb question
They are almost claiming FHE, isn't it just a matter of creating the right tool to get the generated tokens from RAM before it gets encrypted for transfer. How is it fundamentally different than chutes?
Hi! Others have noted this too but you can't seem to buy credits right now, it says "This page couldn’t load" in a custom error page when you've selected the amount and click continue to checkout. Congrats on launching such a cool project that's getting people excited, thinking and discussing :)
Hey Gajesh! I sent you an email with some of the teething problems I ran into trying to get started as a provider. Hope it didn't end up in your spam folder!
Away this looks like a great idea and might have a chance at solving the economic issue with running nodes for cheap inference and getting paid for it.
the MDM profile requirement is suspect though I get why they are doing it. but it doesn't inspire confidence to see that their profile is unsigned and still using the default micromdn scep challenge...
That solution actually makes great sense. So Apple won in some strange way again?
Guess there are limitations on size of the models, but if top-tier models will getting democratized I don’t see a reason not to use this API. The only thing that comes to me is data privacy concerns.
I think batch-evals for non-sensitive data has great PMF here.
Heh, what did they win exactly? This is just a way for another company to extract value out of the single region of the world where Apple is a relevant vendor, and it happens to be the one where it's the easiest to pull people into schemes.
Maybe it's a Baader-Meinhoff phenomena and even a small difference in frequency feels overwhelming because you can't help but notice when it happens without noticing when it doesn't occur.
I have a hard time believing their numbers. If you can pay off a mac mini in 2-4 months, and make $1-2k profit every month after that, why wouldn’t their business model just be buying mac minis?
The numbers are optimistically legit -- it's calculated based purely considering we have demand for all machines at all times. We don't have that right now, but fairly optimistic that people will do it.
That's why we don't recommend purchasing a new machine. Existing machine is no cost for you to run this.
Electricity is one cost, but it will get paid off from every request it receives. Electricity is only deducted when you run an inference. If you have any questions, DM me @gajesh on Twitter.
> That's why we don't recommend purchasing a new machine. Existing machine is no cost for you to run this.
You misunderstood. If the ROI is there, there is enough capital in existence for you to accelerate your profit. So why even deal with the complexity of renting people's hardware when you can do it yourself?
No, what he's saying is that he expects this to be the ROI in the future because his product is so good.
You're not taking into account the thermal strain on the machine, though. A machine that's 100% utilized (even worse if it's in bursts) will last less than an idle machine.
Not appreciably, and not before a 5-yr AppleCare+ warranty expires.
Out of our >3000 currently active Apple Silicon Macs, failures due to non-physical damage are in the single digits per year. Of those, none have been from production systems with 24/7 uptime and continuous high load, which reflects your parenthetical.
Perhaps we haven't met the other end of the bathtub curve yet, but we also won't be retaining any of these very far beyond their warranty period, much less the end of their support life.
AppleCare+ annual is perpetual as long as you keep paying it (and Apple offers to switch to that when your 3-year lump sum expires if you choose that instead). I’m guessing it ends whenever they officially discontinue hardware support, which has traditionally been about 7 years after the last unit is produced, but I haven’t reached that yet to know for sure.
I think the point of this is more "use the machine you have at home" than "do a TCO analysis and see if it's profitable", though. People like to keep their machines working for longer, generally.
> Not appreciably, and not before a 5-yr AppleCare+ warranty expires.
It’s three years for Macs, though I believe you can pay annually for longer. Five has never been a thing to my knowledge.
> A machine that's 100% utilized (even worse if it's in bursts) will last less than an idle machine.
How much though? Say I have three Mac Minis next to each other, one that is completely idle but on, one that bursts 100% CPU every 10 minutes and one that uses 100% CPU all the time, what's the difference on how long the machines survives? Months, years or decades?
> Existing machine is no cost for you to run this.
That is not at all how modern chips work. Idle chips are mostly powered down, non-idle ones are working and that causes real measurable wear and tear on the silicon. CPU, RAM, NAND all wear and tear measurably with use on current manufacturing processes.
https://en.wikipedia.org/wiki/Electromigration
Like pitching "drive rideshares for only the cost of your time & gas"
The question is, do they wear faster than they become obsolete, as in much more expensive to run than buying a new one with higher compute/watt. (and you can also factor in the ability to run latest models at usable speed)
It’s complicated. When you design with modern PDKs, you consider the expected duty cycle of the device, expected temps, and the wear and tear on the silicon all together. That affects the layout of the chip as well as certain choices about widths of various features. Generally, one designs consumer SoCs to last 10 years with the expected duty cycle (low). With more wear you could run out of your “years" much faster, maybe even before the warranty.
I don't worry about bandwidth or constant CPU use, but the one thing that will kill my mac is burning out the SSD.
The calculator gives numbers for nearly everything, but I can't obviously see how much space it needs for model storage or how many writes of temp files I should expect if I'm running flat out.
I assumed you'd want to run one immutable model that can fit in memory without any temp files.
put that stuff on an external disk perhaps, it will eventually crater, but it's easier to replace than macbook internal storage (how are they doing mac minis these days?)
Isn't high memory pressure will result in high memory paging which will wear down the internal ssd?
My question is why did you have to design this to use an MDM instead of a simple program running in the terminal or something?
If you start buying minis, then you need to house, power, and cool them. So you are building a mini data center. If you are building a small data center, economies of scale will drive you to want to build larger and larger. However, this gets expensive and neighbors tend to not like data centers (for good reason). To me this seems like asymmetric warfare against hyper-scalers.
Yup. This way, the people pay for the air conditioning themselves and they probably don't even notice the extra cost.
& if they live in a cool place, they're getting a small space heater as a bonus.
Because they don’t have that much initial money in their pocket, while the idle computer is already there, and the biggest friction point is convincing people to install some software. Both producing rhetoric and software are several order of magnitude cheaper than to directly own and maintain a large fleet of hardware with high guarantee of getting the electrical stable input in a safe place to store them.
Assuming that getting large chunk of initial investment is just a formality is out of touch with 99% of people reality out there, when it’s actually the biggest friction point in any socio-economical endeavour.
No provider maintains 100% utilization of GPUs at full rate. Demand is bursty - even if this project is successful, you might expect, e.g., things to be busy during the stock market times when Claude is throwing API errors and then severely underutilized during the same times that Anthropic was offering two-for-one off peak use.
And then there's a hit for overprovisioning in general. If the network is not overprovisioned somewhat, customers won't be able to get requests handled when they want, and they'll flee. But the more overprovisioned it is, the worse it is for compute seller earnings.
I suspect an optimistic view of earnings from a platform like this would be something like 1/8 utilization on a model like Gemma 4. Their calculator estimates my m4 pro mini could earn about $24/month at 3 hours/day on that model. That seems plausible.
Hold my beer: https://imgur.com/a/sNAoghL
What's your Y axis?
Solid q. I think the part of it is that it’s really easy to attract some “mass” (capital) of users, as there are definitely quite a few of idle Macs in the world.
Non-VC play (not required until you can raise on your own terms!) and clear differentiation.
If you want to go full-business-evaluation, I would be more worried about someone else implementing same thing with more commission (imo 95% and first to market is good enough).
I think the point they’re making though is that the numbers seem too good to be true.
ie. Does anyone know the payback time for a B100 used just for inference? I assume it’s more than a couple of months? Or is it just training that costs so much?
Eigenlayer (which this is spun off from) is a massively VC-funded crypto company.
It is too good to be true. When you see it is making more than a claude code subscription for fuck all work per day.
Prolly gonna make $50 a year tops.
Or like anything else it will be too good to be true at the very beginning but then once people hear about it and it gets popular supply overtakes demand and the mac minis go back to being idle most of the day.
When YouTubers start making videos about it you know it's too late.
The question is what is the hassle of running it plus wear and tear. Max price will tend to that. It is not like crypto where there is capital investment in a rig that can do nothing else. People are using their existing laptop. So I reckon 20-50 a year max per laptop.
The numbers are obviously high, because if this takes off then the price for inference will also drop. But I still think it’s a solid economic model that benefits low income countries the most. In Ukraine, for example, I know people who live on $200/month. A couple Mac Minis could feed a family in many places.
As a business owner, I can think of multiple reasons why a decentralized network is better for me as a business than relying on a hyperscaler inference provider. 1. No dependency on a BigTech provider who can cut me off or change prices at any time. I’m willing to pay a premium for that. 2. I get a residential IP proxy network built-in. AI scrapers pay big money for that. 3. No censorship. 4. Lower latency if inference nodes are located close to me.
How many of those people who could live off $200USD/month can afford or already have a mac mini in the house?
They already have an iPhone. They could save up or borrow for a Mac Mini if they had to. Some of those people I know who live on $200/month have $30k in the bank.
then you are talking about low spend, not low income
Not really. There are lots of people who have low income and low spending, but not low savings. Retired pensioners with savings. Young families who inherited from deceased parents/grandparents. Highly paid professionals on sabbatical. I've met people from all of those categories in Ukraine who live on $200/month.
Isn't this same premise as "lets buy few GPUs to mine crypto and have passive income"? It didn't work very well and it probably won't work now either. If there is money to be made, bigger players will get in there, buy out all mac minis they can, drive price up for regular people and inevitably drive inference price down so you'll be lucky to get initial investment back
No it's not the same premise at all. Crypto doesn't do anything useful for legitimate businesses. AI inference is very useful for legitimate businesses, and so are residential IP proxies for scraping. And by definition, residential IPs cannot be centralized. And as building GPUs becomes more expensive, the existing pool of second hand unused hardware becomes more valuable, not less. The problem with crypto mining is that it quickly becomes unprofitable for small scale deployments. I'm not sure if AI inference would be, especially for the decentralized benefits of lower latency.
It is the same premise, because the person you are responding to is not talking about the moral implications at all, only about the financial / hardware implications.
Running AI inference increases the power draw, and requires certain hardware.
Mining bitcoin increases the power draw, and requires certain hardware.
OP's point thus stands: Bad players will find places to get far cheaper power than the intended audience, and will buy dedicated hardware, at which point the money you can earn to do this will soon drop below the costs for power (for folks like you and me).
Maybe that won't happen, but why won't that happen?
The main problem with crypto is there is no universal need for it. The demand for crypto doesn’t keep increasing as compute gets cheaper. But the demand for AI inference is only growing, and making it cheaper would likely only increase demand. So it’s not a race to the bottom. Sure hyper focused players can earn more at higher margins. But average players can probably still earn decently. Take for example electricity. It can still be profitable for a home in Germany to install balcony solar and make a little money selling back to the grid even though it’s obviously not as efficient as an industrial power plant. Mom and pop AI inference don’t have to be super efficient as long as they serve a universal need - it will be like balcony solar in Europe.
The residential IP proxy point is i think invalidated by their privacy model. I think they aren’t offering up your IP, just your GPU.
On the latency point - your requests are still going through the coordinator of the system here. So on average strictly worse than a large provider.
You - Darkbloom - Operator - Darkbloom - you, vs
You - Provider - you
---
On the censorship point - this is an interesting risk surface for operators. If people are drawn my decentralized model provisioning for its lax censorship, I'm pretty sure they're using it to generate things that I don't want to be liable for.
If anything, I could imagine dumber and stricter brand-safety style censorship on operator machines.
I'm not talking about Darkbloom specifically, but rather this business model in general. I'm sure a future version of Darkbloom could be P2P for better latency. Or their central operator nodes could be geo-balanced. Liability for censorship doesn't matter if it's truly zero trust. Anyway censorship is not my main concern. Low-latency decentralized inference with no US BigTech dependency is a much bigger selling point in Europe.
It's quite funny thinking about a chimpanzee seeing a lot of bananas thinking this could feed my family and then same with humans only with Mac Minis
> These are estimates only. We do not guarantee any specific utilization or earnings. Actual earnings depend on network demand, model popularity, your provider reputation score, and how many other providers are serving the same model.
Others are reporting low demand, eg.: https://news.ycombinator.com/item?id=47789171
Of course these numbers are ridiculous. Mac Mini (let's assume Apple releases M5 Pro) tops Int8 (let's assume it is the same as FP8, which it is not) at ~50 TFLOPs, with Draw Things, we recently developed hybrid NAX + ANE inference, which can get you ~70 TFLOPs.
A H200 gives you ~4 PFLOPs, which is ~60x at only ~40x price (assuming you can get a Mac Mini at $1000). (Not to mention, BTW, RTX PRO 6000 is ~7x price for ~40x more FLOPs).
Your M4 Mac Mini only has ~20 TFLOPs.
> Your computer only has ~20 TFLOPs.
What a time to be alive.
Power and racking are difficult and expensive?
How difficult? Is running 1000 minis worth $1,000,000/month of effort? I feel like it is.
And at that scale (1k) it ain't even that hard, a single room could be enough to hazardly drop them on shelves with a big fan to draw out the heat
There are many people who do not have ready access to a million dollars to purchase said Mac minis, much less the operating capital to rack & operate them.
Very smart play to build a platform, get scale, and prove out the software. Then either add a small network fee (this could be on money movement on/off platform), add a higher tier of service for money, and/or just use the proof points to go get access to capital and become an operator in your own pool.
If those numbers are true, they could tart with one Mac and can double every few months. But, I guess there are also many people who do not have ready access to whatever a Mac mini costs either...
You can run the simulation out, but if the idea works, you can get to scaled revenue much faster than organic growth keeping 100% of the margin.
This is essentially the same reason even the best money managers take outside money to start, even if they eventually kick out the investors.
Because the "ship software to people, rent their hardware" model has zero up front investment required, presumably. And they don't have to deal with power, cooling, real estate.
"You could see a single robotaxi being worth, or providing, about $30,000 of gross profit per year. ... A Tesla is an appreciating asset..."
- Elon Musk during Tesla's Autonomy Day in April 2019.
Capital and availability?
I guess if it only works at scale capital is maybe the answer. Like enough cash to buy 5 or 10 or even 100 minis seem doable - but if the idea only works well when you have 10,000 running - that makes some sense.
Being the middleman is often way more profitable
Because their numbers don’t work out. When you do the math on token cost versus inference speed, you get something that barely breaks even even with cheap power.
Also they’ve already launched a crypto token, which is a terrible sign.
I installed this so you don't have to. It did feel a bit quirky and not super polished. Fails to download the image model. The audio/tts model fails to load.
In 15 minutes of serving Gemma, I got precisely zero actual inference requests, and a bunch of health checks and two attestations.
At the moment they don't have enough sustained demand to justify the earning estimates.
They released this like a day ago, I'm not surprised that there's not enough demand right now. Give it some time to take off
You'd think to bootstrap a marketplace you'd spend your own money to feed fake requests (or perhaps allow free chat so that they induce requests).
Still, absolute zero is an unacceptable number. Had this running for more than an hour.
I kind of see your point, but I also kind of don't.
Sure, it would be great if you'd immediately get hammered with hundreds of requests and start make money quickly. It would also be great if it was a bit more transparent, and you could see more stats (what counts as "idle"? Is my machine currently eligible to serve models?). But it's still very new, I'd say give it some time and let's see how it goes.
If you have it running and you get zero requests, it uses close to zero power above what your computer uses anyway. It doesn't cost you anything to have it running, and if you get requests, you make money. Seems like an easy decision to me.
Bootstrapping will be near-impossible (or incredibly costly) unless they offer inference consumers models with established demand arriving at some least-cost router service where they can undercut the competition (if they actually can). And then dogfood the opportunistic provider side on their own Macs, but with a preference to putting third parties first in the queue. Everything else is just wishful thinking.
Well I already made the Ctrl+C decision. Yours may have been different, but I suppose only one of us installed it, and that one counts.
I went with the ctrl z approach.
Hopefully you also set it running in the background.
Copy?
SIGINT
> We're not taking funds from customers yet — we're personally paying for all the provider requests during this phase. Credit purchases are disabled.
This appears on their credit purchase page right now, but you have to email them to get credits (everyone starts with zero)
Has anyone tested the system from the other end... sending a prompt and getting a response?
weird to learn that they do not generate inference requests to their network themselves to motivate early adopters at least to host their inference software
If they paid promised > $1k/m for FLUX 2B on a Mac they would go broke in less than a month. On a single 5090 that model would provide an inference througput so high they'd have to pay close to $50k/m for the results.
The numbers are absolute fraud. You shouldn't be installing their software cause fraud could be not just about numbers.
Can you rephrase that? I don't think I've read it correctly. It sounds like you are saying it would normally cost $50k on a 5090 and they can do equivalent work paying $1k. That's sounds like a $49k profit margin, but you say they will go broke.
I'm assuming it's meant the other way around.
Given their estimates of a Mac being able to generate $1k (per month?) a 5090 with a lot more power would be able to generate $50k. For a $3k piece of hardware. Which is obviously not realistic. (As in, nobody is paying that much for the images, which seems to match well with no actual requests on the system.)
and I don't think they ever will unless they're highly competitive (hopefully that price they have stays? at least for users)
I was thinking of building this exact thing a year ago but my main stopper was economics: it would never make sense for someone to use the API, thus nobody can make money off of zero demand.
I guess we just have to look at how Uber and Airbnb bootstrapped themselves. Another issue with my original idea was that it was for compute in general, when the main, best use-case, is long(er)-running software like AI training (but I guess inference is long running enough).
But there already exist software out there that lets you rent out your GPU so...
People underestimate how efficient cost/token is for beefy GPUs if you are able to batch. Unlikely for one off consumer unit to be able to compete long term.
What's a good place to do this?
For Windows there's https://borg.games/setup (I'm the author).
How much revenue to users actually see?
[dead]
You can see in their stats view they have a lot of providers/nodes connected but practically no actual demand/consumers. They just launched and I'm sure get providers was top of their agenda, but it's essentially unusable as a provider unless they perform some serious lift to get actual paying customers.
I received the same error, but it was followed by this line in the logs, which might explain the lack of inference requests assume there is actual demand...
WARN STT backend failed health check — model will NOT be advertised
I disabled the audio model (had to remove it, it was so buggy) and then started up with a text only model. Serving started without error. The system simply has no requests. The economics seem like a mirage anyway.
You have to install their MDM device management software on your computer. Basically that computer is theirs now. So don't plan on just handing over your laptop temporarily unless you don't mind some company completely owning your box. Still might be a validate use for people with slightly old laptops lying around, but beware trying to share this computer with your daily activities if you e.g. use a bank on a browser on this computer regularly. MDM means they can swap out your SSL certs level of computer access, please correct me if I'm wrong.
MDMs on macOS are permissioned via AccessRights, and you can verify that their permission set is fairly minimal and does not allow what you've described here (bits 0, 4, 10).
That said, their privacy posture at the cornerstone of their claims is snake oil and has gaping holes in it, so I still wouldn't trust it, but it's worth being accurate about how exactly they're messing up.
Edit: deleted post. I see your other post now.
You are right - the "nonce binding" the paper uses doesn't seem convincing. The missing link is that Apple's attestation doesn't bind app generated keys to a designated requirement, which would be required to create a full remote attestation.
> If you can prove a public key is generated by the SEP of a machine running with all Apple's security systems enabled, then you can trivially extend that to confidential computing because the macOS security architecture allows apps to block external inspection even by the root user.
It only effectively allows this for applications that are in the set of things covered by SIP, but not for any third-party application. There's nothing that will allow you to attest that arbitrary third-party code is running some specific version without being tampered with, you can only attest that the base OS/kernel have not been tampered with. In their specific case, they attempt to patch over that by taking the hash of the binary, but you can simply patch it before it starts.
To do this properly requires a TEE to be available to third-party code for attestation. That's not a thing on macOS today.
I wiped my post because you are right. I don't think it needs a full SGX-style TEE. What's missing is a link to designated requirements. Abusing a nonce field doesn't seem to work, or if it does I can't figure out how. The MDM/MDA infrastructure would need to be able to include:
The macOS KeyStore infrastructure does track this which is why I thought it'd work. But the paper doesn't mention being able to get this data server side anywhere. Instead there's this nonce hack.It's odd that the paper considers so many angles including things like RDMA over Thunderbolt, but not the binding between platform key and app key.
Reading the paper again carefully I get the feeling the author knows or believes something that isn't fully elaborated in the text. He recognizes that this linkage problem exists, proposes a solution and offers a security argument for it. I just can't understand the argument. It appears APNS plays a role (apple push notification service) and maybe this is where app binding happens but the author seems to assume a fluency in Apple infrastructure that I currently lack.
I can buy the idea that if you can have the MDM infrastructure attest the code signing identity through the designated requirements, that you can probably come pretty close, but I'm still not quite sure you get there with root on macOS (and I suspect that this is part of why DCAppAttest hasn't made it to macOS yet).
Certainly, it still doesn't get you there with their current implementation, as the attempts at blocking the debugger like PT_DENY_ATTACH are runtime syscalls, so you've got a race window where you can attach still. Maybe it gets you there with hardened runtime? I'd have to think a bit harder on that.
Yeah I didn't quite understand the need for PT_DENY_ATTACH. Hardened runtime apps that don't include get-task-allow are already protected from debugger attach from the start of the process, unless I misunderstood something.
I'm not quite sure why Apple haven't enabled DCAppAttest on macOS. From my understanding of the architecture, they have every piece needed. It's possible that they just don't trust the Mac platform enough to sign off on assertions about it, because it's a lot more open so it's harder to defend. And perhaps they feel the reputational risk isn't worth it, as people would generalize from a break of App Attest on macOS to App Attest on iOS where the money is. Hard to say.
MDM is the absolute deal breaker. No way in hell will I ever make my Macbook into an unsellable brick if they so decide to lock my computer with MDM.
Even moreso, not for pennies/month
They can never have supervised permissions on your device unless your mac was wiped and registered in Apple Business Manager. As a BYOD, it cannot be locked.
Whether it makes sense to enroll a mac as a BYOD is another question.
The MDM profile doesn’t grant that
Unfortunately, verifiable privacy is not physically possible on MacBooks of today. Don't let a nice presentation fool you.
Apple Silicon has a Secure Enclave, but not a public SGX/TDX/SEV-style enclave for arbitrary code, so these claims are about OS hardening, not verifiable confidential execution.
It would be nice if it were possible. There's a lot of cool innovations possible beyond privacy.
I wrote a whole SDK for using SGX, it's cool tech. But in theory on Apple platforms you can get a long way without it. iOS already offers this capability and it works OK.
macOS has a strong enough security architecture that something like Darkbloom would have at least some credibility if there was a way to remotely attest a Mac's boot sequence and TCC configuration combined with key-to-DR binding. The OS sandbox can keep apps properly separated if the kernel is correct and unhacked. And Apple's systems are full of mitigations and roadblocks to simple exploitation. Would it be as good as a consumer SGX enclave? Not architecturally, but the usability is higher.
As if you get privacy with the inference providers available today? I have more trust in a randomly selected machine on a decentralized network not being compromised than in a centralized provider like OpenAI pinky promising not to read your chats.
Inference providers don't claim private inference. However, they must uphold certain security and legal compliances.
You have no guarantees over any random connected laptop connected across the world.
I would say the chances of OpenAI itself getting hacked and your secrets in logs getting leaked are about the same or less as the chances of a randomly selected machine on a decentralized network being reverse-engineered by a determined hacker. There's no risk-free option, every provider comes with risks. If you care about infosec you have to do frequent secret rotation anyway.
Every hardware key will be broken if there is enough incentive to do so. Their claims read like pure hubris.
Who cares about AI privacy? Most people don’t. If you do, run locally.
They use the TEE to check that the model and code is untampered with. That's a good, valid approach and should work (I've done similar things on AWS with their TEE)
The key question here is how they avoid the outside computer being able to view the memory of the internal process:
> An in-process inference design that embeds the in- ference engine directly in a hardened process, elimi- nating all inter-process communication channels that could be observed, with optional hypervisor mem- ory isolation that extends protection from software- enforced to hardware-enforced via ARM Stage 2 page tables at zero performance cost.[1]
I was under the impression this wasn't possible if you are using the GPU. I could be misled on this though.
[1] https://github.com/Layr-Labs/d-inference/blob/master/papers/...
This entire paper smells of LLM, I'm sure even the most distinguished academic would refrain from using notation to prove that the SIP status cannot change during operation.
While they do make this argument, realistically anyone sending their prompt/data to an external server should assume there will be some level of retention.
And more so in particular, anyone using Darkbloom with commercial intents should only really send non-sensitive data (no tokens, customer data, ...) I'd say only classification tasks, imagine generation, etc.
There's a difference between trusting Anthropic and trusting random mac owners.
I know where my answer lies in that; but i don't claim to be an objective truth.
For example OpenAI has been caught sharing data with the gov. agencies.
Apple Silicon systems have unified memory between CPU and GPU. The hypervisor page table trick is thus claimed to protect GPU memory from RDMA.
Macs do not have an accessible hardware TEE.
Macs have secure enclaves.
Good point!
But they argue that:
> PT_DENY_ATTACH (ptrace constant 31): Invoked at process startup before any sensitive data is loaded. Instructs the macOS kernel to permanently deny all ptracerequests against this process, including from root. This blocks lldb, dtrace, and Instruments.
> Hardened Runtime: The binary is code-signed with hardened runtime options and explicitly without the com.apple.security.get-task-allow entitlement. The kernel denies task_for_pid() and mach_vm_read()from any external process.
> System Integrity Protection (SIP): Enforces both of the above at the kernel level. With SIP enabled, root cannot circumvent Hardened Runtime protections, load unsigned kernel extensions, or modify protected sys- tem binaries. Section 5.1 proves that SIP, once verified, is immutable for the process lifetime.
gives them memory protection.
To me that is surprising.
Looking at their paper at [1], there's a gaping hole: there's no actual way to verify the contents of the running binaries. The binary hash they include in their signatures is self-reported, and can be modified. That's simply game over.
[1] https://github.com/Layr-Labs/d-inference/blob/master/papers/...
A note, as others have posted on this thread: I mention this as a concrete and trivial flaw in their whole strategy, but the issue is fundamental: there's no hardware enclave for third-party code available to do the type of attestation that would be necessary. Any software approach they develop will ultimately fall to that hole.
Couldn't someone just uhh... patch their macOS/kernel, mock these things out, then behold, you can now access all the data?
If it's not running fully end to end in some secure enclave, then it's always just a best effort thing. Good marketing though.
Right.
Apple is perfectly capable of doing remote attestation properly. iOS has DCAppAttest which does everything needed. Unfortunately, it's never been brought to macOS, as far as I know. Maybe this MDM hack is a back door to get RA capabilities, if so it'd certainly be intriguing, but if not as far as I know there's no way to get a Mac to cough up a cryptographic assertion that it's running a genuine macOS kernel/boot firmware/disk image/kernel args, etc.
It's a pity because there's a lot of unique and interesting apps that'd become possible if Apple did this. Darkbloom is just one example of what's possible. It'd be a huge boon to decentralization efforts if Apple activated this, and all the pipework is laid already so it's really a pity they don't go the extra mile here.
> iOS has DCAppAttest which does everything needed. Unfortunately, it's never been brought to macOS, as far as I know.
Apple's docs claim it's been available on macOS since macOS 11. Am I missing something here?
https://developer.apple.com/documentation/devicecheck/dcappa...
All lies. They mean the symbols exist and can be linked against, but
https://developer.apple.com/documentation/devicecheck/dcappa...
> If you read isSupported from an app running on a Mac device, the value is false. This includes Mac Catalyst apps, and iOS or iPadOS apps running on Apple silicon.
That really sucks! TIL. So app attestation is iOS 14.0+, iPadOS 14.0+, tvOS 15.0+ and watchOS 9 only.
Yes. Running attested workloads on macOS if you are not Apple is nontrivial.
[dead]
You can probably just tap the HTTP(S) connection to spy on the data coming through. I think it's a mistake to assume any kind of privacy for this service.
The biggest argument for remote attestation I can think of is to make sure nobody is returning random bullshit and cashing in prompt money on a massive scale.
> PT_DENY_ATTACH
All you have to do is attach to the process before it does that, and then prevent this call from going through.
They quite frankly have no idea what they are talking about.
I'm not arguing anything. This is how it works. There is no but.
Protection here is conditional, best-effort. There are no true guarantees, nor actual verifiability.
Cool idea. Just some back-of-the-envelope math here (not trusting what's on their site):
My M5 Pro can generate 130 tok/s (4 streams) on Gemma 4 26B. Darkbloom's pricing is $0.20 per Mtok output.
That's about $2.24/day or $67/mo revenue if it's fully utilized 24/7.
Now assuming 50W sustained load, that's about 36 kWh/mo, at ~$.25/kWh approx. $9/mo in costs.
Could be good for lunch money every once in a while! Around $700/yr.
Well. Running your machine to do inference will utilize more than 50W sustained load, I'd say more than double that. Plus electricity is more expensive here (but granted, I do have solar panels). Plus don't forget to factor in that your hardware will age faster.
I'd say it's not worth it. But the idea is cool.
Your hardware will age slower if you have consistent load.
Thermal stress from bursty workloads is much more of a wearing problem than electromigration. If you can consistently keep the SoC at a specific temperature, it'll last much longer.
This is also why it was very ironic that crypto miner GPUs would get sold at massive discounts. Everyone assumed that they had been ran ragged, but a proper miner would have undervolted the card and ran it at consistent utilization, meaning the card would be in better condition than a secondhand gamer GPU that would have constantly been shifting between 1% to 80% utilization, or rather, 30°C to 75°C
Their estimate is based on significantly lower consumption when under load. E.g. 25W for an M4 Pro mac mini. I have no idea if that’s realistic - but the m4s are supposedly pretty efficient (https://www.jeffgeerling.com/blog/2024/m4-mac-minis-efficien...)
Their example big earner models are FLUX.2 Klein 4B and FLUX.2 Klein 9B, which i imagine could generate a lot more tokens/s than a 26B model on your machine.
For Gemma 4 26B their math is:
single_tok/s = (307 GB/s / 4 GB) * 0.60 = 46.0 tok/s
batched_tok/s = 46.0 * 10 * 0.9 = 414.4 tok/s
tok/hr = 414.4 * 3600 = 1,492,020
revenue/hr = (1,492,020 / 1M) * $0.200000 = $0.2984
I have no idea if that is a good estimate of how much an M5 Pro can generate - but that’s what it says on their site.
They do a bit of a sneaky thing with power calculation: they subtract 12Ws of idle power, because they are assuming your machine is idling 24/7, so the only cost is the extra 18W they estimate you’ll use doing inference. Idk about you, but i do turn my machine off when i am not using it.
Interesting token numbers they're using, because I've benchmarked it at 69 tok/s single steam and 130 multi stream.
> My M5 Pro can generate 130 tok/s (4 streams) on Gemma 4 26B.
This seems high. At which quantization? Using LM Studio or something else?
Note: Darkbloom seems to run everything on Q8 MLX.
Ah good point, this is using Q4, benchmarked total throughout serving with Llama.cpp.
OpenAI has only about 5% paying customers, how does it generate revenue?
I don’t think this is a sustainable business model. For example, Cubbit tried to build decentralised storage, but I backed out because better alternatives now exist, and hardware continues to improve and become cheaper over time.
Your electricity and ownership are going to get lower return and does not actually requce CO2.
Also this assumes hardware never fails. I learned about this the hard way back when I started mining crypto on my 5700XT way back when.
I figured since I already used it a lot, and I've never had a GPU fail on me, it would be fine.
The fans on it died in a month of constant use, replacing them was more money than what I made on mining.
Genuinely curious, is there any way to estimate amortization of Mac?
I’d imagine 1 year of heavy usage would somehow affect its quality.
Yeah, only way to get there is assuming they're not giving prompt caching discounts while my laptop is getting prompt caching benefits, with very many large prompts. So yes I am skeptical of their numbers.
Any idea what makes for such a diff between your and theirs numbers? Batching? Or could they do a crazy prefix caching across all nodes to reduce the actual processing.
Maybe lunch money for you, but there are people in some parts of the world who live on $200/month. Like Ukraine.
But they probably don’t have M5 MacBook pros idling
Or reliable energy or internet.
They can acquire one if it offers real opportunities like this.
Don't forget to factor in cooling costs.
Or saved heating costs in the winter!
You're printing everyone's serial numbers publicly. https://console.darkbloom.dev/providers then "Security Verification" for any machine and then "Verify this device independently" -- all of this can be scraped.
This is one of those ideas I think makes perfect sense, but requires so much operational change for the entire stack, that it would be very difficult to scale:
- Convincing labs to run distributed, burst-y inference
- Convincing people to run their Mac all day, hoping to make a little profit
- Convincing users to trust a distributed network of un-trusted devices
I had a similar idea, pre-AI, just for compute in general. But solving even 1 of those 3 (swap AI lab for managed-compute-type-company, eg Supabase, Vercel) is nearly impossible.
...or convincing operators that jobs sent to their machines are legal, legitimate, and non-nefarious.
I could not find disclosure on their site about the guard-railing or safety-systems at the point the prompt is gathered from users which would intercept, log & prevent bad actors from inadvertently involving me in something illegal or immoral as an operator. Perhaps that disclosure exists and I just need to be linked to it; that would be welcome.
Having strong SETI@Home vibes from 25 years ago, except of course, this is not for the greater good of humanity, but a for-profit project.
Problem is, from a technical point of view, what kind of made sense back then (most people running desktops, fans always on, energy saving minimal) is kind of stupid today (even if your laptop has no fan, would you want it to be always generating heat?)...
I definitely want my laptops to be cool, quiet and idle most of the time.
My m4 max mbp with 128gb of memory is constantly training 24/7 on weekends- it’s why I bought the thing.
I some times play about with local models via ollama/comfyui and more recently ace-step to generate music.
This is short bursts of heat 5-10 m during the render I would not be happy with that for multiple hours a day. I am sure that would have a negative effect on battery health.
@eigengajesh - Your cost estimator lists Mac Mini M4 Pro with only 24 or 48GB options, but the M4 Pro mini can also be configured with 64GB. At least, I hope so, as I'm typing this on one. ;-)
Oh, also, you seem to have some bugs:
Gemma: WARN [vllm_mlx] RuntimeError: Failed to load the default metallib. This library is using language version 4.0 which is not supported on this OS. library not found library not found library not found
cohere: 2026-04-16T14:25:10.541562Z WARN [stt] File "/Users/dga/.darkbloom/bin/stt_server.py", line 332, in load_model 2026-04-16T14:25:10.541614Z WARN [stt] from mlx_audio.stt.models.cohere_asr import audio as audio_mod 2026-04-16T14:25:10.541643Z WARN [stt] ModuleNotFoundError: No module named 'mlx_audio.stt.models.cohere_asr'
Trying to download the flux image models fails with:
curl: (56) The requested URL returned error: 404
darkbloom earnings does not work
your documentation is inconstent between saying 100% of revenue to providers vs 95%
I think .. this needs a little more care and feeding before you open it up widely. :) And maybe lay off the LLM generated text before it gets you in trouble for promising things you're not delivering.
[dead]
Had it for 3 days running with nary a request. Perhaps it's the chosen model that I am serving (Gemma 4 26b)? I did see some WARN logs right after startup but the description don't suggest any problems that would block accepting requests or processing them.
I'd love a way to do this locally -- pool all the PCs in our own office for in-office pools of compute. Any suggestions from anyone? We currently run ollama but manually manage the pools
https://github.com/exo-explore/exo
If you set CPUSchedulingPolicy=idle Nice=19 IOSchedulingClass=idle in the ollama server configuration it should run in the background with lowest priority.
Seems like so much more work than "just" paying for https://huggingface.co or whichever other neocloud who already did all the setup for you and just waits for your credit card per minute/seconds/token.
It is much more work because for many workloads you have geographic ringfencing and cannot send it out to the cloud
Doubt this kind of workloads would agree to send data then to a cloud of randos devices, precisely when cloud providers to certify they aren't looking at clients data (Customer-managed encryption keys, CMEK).
>> Doubt this kind of workloads would agree to send data then to a cloud of randos devices,
Totally agree, which is why i said "I'd love a way to do this locally -- pool all the PCs in our own office for in-office pools of compute."
Interesting concept. Two sided marketplaces are hard to bootstrap but maybe just enough curiosity would get the flywheel going. Hell they should just try and convince people to enroll as providers but then also use the service even if it’s hitting their own machines until there’s some degree of supply and demand pressure then try and get only providers to sign up. Or set up some way to encourage providers to promote others to use the service (the 100% rev share kind of breaks that concept but anything can change).
I wish this was self hostable, even for a license fee. Many businesses have fleets of Macs, sometimes even in stock as returned equipment from employees. Would allow for a distributed internal inference network, which has appeal for many orgs who value or require privacy.
So basically ... Pied Piper.
finally!
You might not even know it as a user but the payment/distribution here is all built on crypto+stablecoins. This is a great use case for it.
Good. Another great non-speculative use-case for crypto and stablecoins.
Amazing! Let me see, doing the math r/n… carry the one, yup that makes the total number of non-speculative uses for crypto and stablecoin: 1
;P
It always has been payments. x402 and Stripe Tempo makes the use case more than 1.
I'm not sure how the economics works out. Pricing for AI inference is based on supply/demand/scarcity. If your hardware is scarce, that means low supply; combine with high demand, it's now valuable. But what happens if you enable every spare Mac on the planet to join the game? Now your supply is high, which means now it's less valuable. So if this becomes really popular, you don't make much money. But if it doesn't become somewhat popular, you don't get any requests, and don't make money. The only way they could ensure a good return would be to first make it popular, then artificially lower the number of hosts.
Generate images requested by randoms on the internet on your hardware.
What could possibly go wrong?
Love the concept, with some similarity to folding@home, though more personal gain.
But trying it out it still needs work, I couldn't download a model successfully (and their list of nodes at https://console.darkbloom.dev/providers suggests this is typical).
And as a cursory user, it took me some digging to find out that to cash out you need a Solana address (providers > earnings).
"Debugger attachment is blocked. Memory inspection is blocked." - reminds me old crackme challenges. Everything they mention can be bypassed, so determined person can start stealing data from the network. For me this is a killer of such distributed compute ideas, but who knows, maybe the non-enteprise users will be desperate enough for cheap compute to make this idea valid.
The hardware-attested privacy path is the interesting part of this, but the economic side has a quieter risk the thread has not named: the load tax per request. MiniMax M2.5 239B from your catalog still has to load all 239B weights even though only 11B are active — that is roughly 120GB at Q4_K_M, and cold load from SSD on Apple Silicon is measurable in tens of seconds. Even the Qwen3.5 122B MoE lands around 65GB cold. If the coordinator routes request number two to a different idle Mac than request number one, or if the owner's machine spun the model out to free memory in between, each request pays that cold load before the first token. Keeping the model resident 24/7 solves the latency but eats into the power budget the operator is trying to amortize in the first place. How does the coordinator decide which provider to keep warm for which model? A 16GB or 32GB home Mac cannot host Qwen3.5 122B MoE at all, and the Mac Studios that can are a much smaller slice of the 100M machine estimate.
I like the idea but it wont take off until Homomorphic Encryption for inference becomes a thing that's efficient and anyone can be a node.
This is what I was confused about like if they own their device they can see what's happening there until we solve this
Interesting to see an offering with this heritage [1] proposing flat earnings rates for inference operators here, rather than trying to sell a dynamic marketplace where operators compete on price in real-time.
Right now the dashboards show 78 providers online, but someone in-thread here said that they spun one up and got no requests. Surely someone would be willing to beat the posted rate and swallow up the demand?
I expect this is a migration target, but a tactical omission from V1 comms both for legitimate legibility reasons (I can sell x for y is easier to parse than 'I can participate in a marketplace') and slightly illegitimate legibility reasons (obscuring likely future price collapse).
Still - neat project that I hope does well.
[1] Layer Labs, formerly EigenLayer, is company built around a protocol to abstract and recycle economic security guarantees from Ethereum proof of stake.
[dead]
I won't install some random untrusted binary off of some website. I downloaded it and did some cursory analysis instead.
Got the latest v0.3.8 version from the list here: https://api.darkbloom.dev/v1/releases/latest
Three binaries and a Python file: darkbloom (Rust)
eigeninference-enclave (Swift)
ffmpeg (from Homebrew, lol)
stt_server.py (a simple FastAPI speech-to-text server using mlx_audio).
The good parts: All three binaries are signed with a valid Apple Developer ID and have Hardened runtime enabled.
Bad parts: Binaries aren't notarized. Enrolls the device for remote MDM using micromdm. Downloads and installs a complete Python runtime from Cloudflare R2 (Supply chain risk). PT_DENY_ATTACH to make debugging harder. Collects device serial numbers.
TL;DR: No, not touching that.
[fix: remove hardcoded API_KEYS ](https://github.com/Layr-Labs/d-inference/pull/39/changes)
oh boooy, it's a benchmarking script, but still...
Until we have breakthroughs in homomorphic encryption compute, I won't trust such privacy claims
Why only Macs? If we think of all PCs and mobile phones running idle, the potential is much larger.
From the paper: https://github.com/Layr-Labs/d-inference/blob/master/papers/...
> Apple’s attestation servers will only generate the FreshnessCode for a genuine device that checks in via APNs. A software-only adversary cannot forge the MDA certificate chain (Assumption 3). Com- bined with SIP enforcement (preventing binary replace- ment) and Secure Boot (preventing bootloader tampering), this provides strong evidence that the signing key resides in genuine Apple hardware.
I am not entirely sure they understand that System Integrity Protection and Secure Boot can be turned off.
My understanding from the paper is that doing so should cause certain things in Apple's hardware security enclaves to break a signing chain, and a server-side MDM system integrated with Apple servers can detect this. But I'm not familiar with the underlying technology, so not sure if underlying assumptions are incorrect.
AFAIK that just ensures the SEP is present but perhaps they are signing the boot state now
They use the Apple TEE which they claim also protects GPU memory (I wasn't aware of this).
NVidia data center GPUs have a similar path, but not their consumer ones. Not sure about the NVidia Spark.
It's possible AMD Strix Halo can do this, but unlikely for any other PC based GPU environments.
Epyc has that VM encrypted memory thing, which comes pretty close. It does raise an interesting question, though: would a PCIe card passed through to a VM be able to DMA access the memory of neighboring devices?
simple first target, PCs have more variability
"These are estimates only. We do not guarantee any specific utilization or earnings. Actual earnings depend on network demand, model popularity, your provider reputation score, and how many other providers are serving the same model.
When your Mac is idle (no inference requests), it consumes minimal power — you don't lose significant money waiting for requests. The electricity costs shown only apply during active inference.
Text models typically see the highest and most consistent demand. Image generation and transcription requests are bursty — high volume during peaks, quiet otherwise."
It's a good project that makes sense. I recommend adding a contractual layer as well, since it's free and makes sense. Operators could legally sign that they will not look into the inference layer. After all, the operators already have a financial relationship with this provider, so it makes sense to add a contract to it and keep operators from looking into other people's data that way, too. I wish this project a lot of success.
Wasn't there an idea about 15 years ago where you would open your browser, go to a webpage and that page would have a JavaScript based client that would run distributed workloads?
I believe the idea was that people could submit big workloads, the server would slice them up and then have the clients download and run a small slice. You as the computer owner would then get some payout.
Intersting to see this coming back again.
I used to work at Distributive (formerly "Kings Distributed Systems") on its DCP compute platform" which is entirely what you're describing. You can deploy a JS/WASM based workload, and it will be "sliced" and served to browser-based compute nodes. With WebGPU you can sort of have inference executing in the browser too. Incredible people there with an awesome project
I added Python execution support via Pyodide (cpython compiled to wasm) and worked on a bunch of other random stuff like WebLLM inferencing during my time there.
Apart from Distributive, there's also the "Golem network", "Salad", "Koii" and various other similar projects.
---
I'm not sure if I'm convinced by the "Uber for compute" use case with compute buyers and compute workers (sellers), but if you're a university and you have 1000 Windows machines across all your computer labs, it'd be nice to leverage that compute for running research or something idk - especially with the price of ram / cloud offerings these days...
> but if you're a university and you have 1000 Windows machines across all your computer labs, it'd be nice to leverage that compute for running research or something idk - especially with the price of ram / cloud offerings these days...
This reminds me of the DevOps guy who made the developer laptops part of a Jenkins "swarm" under the thought that the machines were beefy and underutilized most of the time.
Or SETI which would search for signs of alien life.
I think it’s important that systems like this exist, but getting them off the ground is non-trivial.
We’ve been building something similar for image/video models for the past few months, and it’s made me think distribution might be the real bottleneck.
It’s proving difficult to get enough early usage to reach the point where the system becomes more interesting on its own.
Curious how others have approached that bootstrap problem. Thanks in advance.
As one of the only people running a Mac Studio M3 Ultra with 512 GB of RAM on the network, I can tell you at sustained 100% GPU utilization I am measuring 250 watts max (at the power outlet). My solar panels are easily producing this. The power calculation goes away once you connect a solar panel. You can get a 400 watt solar panel on Amazon for $300.
> You can get a 400 watt solar panel on Amazon for $300.
Too expensive. It's probably producing 200 watts average for 8 hours a day. That's 1600 watt hours, which is about $1.60 at PG&E prices. That would take 187 days to recoup the cost of just the panel.
If you include installation costs and "what PG&E steals if you wire it to the same grid" it's probably more like 4x that, which is too long.
Tell me when we can have 400 watt solar panels for $50. Stupid capitalism literally forces solar panel prices to make it unprofitable.
People should never have to take out loans for solar. Solar should be subsidized and forced by the government to be so cheap that it repays for its cost within a month. Then we're talking. Most things I buy to save money, I expect them to repay within a month. Maybe 2 months max.
You can get them used for that price, and new for $107 if you buy qty 10+. See signature solar as one example.
Installation costs and inverters not included, however.
If you are worried about a $300 solar panel you are not going to like the cost of a Mac Studio M3 Ultra 512 GB! haha
I'm not "worried" about that cost, I would rather just pay PG&E electricity if the solar panel cost $300.
Just pointing out why capitalism + solar is a failure. Capitalism reprices the good thing to be equally expensive to the bad thing, so that nobody buys the good thing anymore.
Are you actually seeing demand?
Like the concept. This is not a business - should be an open source GitHub repo maybe.
They lost me with just one microcopy - “start earning”. Huge red signal.
But why would I donate my Mac Studio's idle time if I couldn't "start earning"?
> Operators cannot observe inference data.
Is there some actual cryptography behind this, or just fundamentally-breakable DRM and vibes?
Reminds me a lot of when I used to deploy this (DataseamGrid) on K12 computers. I was actually just discussing this scenario with a friend.
https://www.dataseam.org/research/
I installed two models, but it just always reports:
Also the benchmark just doesn't work.Interesting idea, but needs some work.
client side of this kind of needs to be open source unless I'm running it on a dedicated machine and firewalling it from the rest of my network. Or the company needs to have a very strong reputation and certifications. curlbash and go is a pretty hard sell for me
Cool idea, though hats off to anyone who got cohere-transcribe to show up as serving the model. I could get device to show up, but kept having issues getting their server to properly serve the model though it could just be the device I tested.
Yeah I think there's a dependency issue going on there. Something isn't installed that needs to be.
How can one do this safely? If I create a new, non-sudo user, can I install the MDM profile only for that user? I don't understand how this all works obviously so maybe this is a very dumb question
They are almost claiming FHE, isn't it just a matter of creating the right tool to get the generated tokens from RAM before it gets encrypted for transfer. How is it fundamentally different than chutes?
hey guys! i'm the creator. let me know if you have any questions.
Hi! Others have noted this too but you can't seem to buy credits right now, it says "This page couldn’t load" in a custom error page when you've selected the amount and click continue to checkout. Congrats on launching such a cool project that's getting people excited, thinking and discussing :)
Hey Gajesh! I sent you an email with some of the teething problems I ran into trying to get started as a provider. Hope it didn't end up in your spam folder!
Please offer new clients try it: at least let us to send few requests in the chat.
Is this named after the 2011 split album with Grimes and d'Eon?
Should have called it “Inferanet” with this idea.
Away this looks like a great idea and might have a chance at solving the economic issue with running nodes for cheap inference and getting paid for it.
I've tried to install it on my mac, but not sure what macOS version it should support.
on 15.1 it failed to serve models.
updated to latest 15.5 and it fails to run binary.
I think macOS has jumped to 26, right?
They could consider registering as a provider on something like OpenRouter if they aren't getting enough inference requests on their own site.
[dead]
Seems like an interesting way for those people that purchased a Mac Mini to run OpenClaw to pay off the hardware, since mostly it’s now idle.
> Every request is end-to-end encrypted
Afaik you will need to decrypt the data the moment it needs to be fed into the model.
How do they do this then?
The system hosting the model must be one of the ends.
Remember, all encryption is E2EE if you're not picky about the ends.
I cant buy credits - says page could not load
Anyone know why not est earnings are shown for Mac Airs?
Also they don’t list the M5 air.
Thanks, if this takse off. I have finally some motivation to do exploitation in kernel. :)
Why does M1 Max project significantly higher revenue than M3 Max with double the ram?
the MDM profile requirement is suspect though I get why they are doing it. but it doesn't inspire confidence to see that their profile is unsigned and still using the default micromdn scep challenge...
This feels like defi... de-ai
That solution actually makes great sense. So Apple won in some strange way again?
Guess there are limitations on size of the models, but if top-tier models will getting democratized I don’t see a reason not to use this API. The only thing that comes to me is data privacy concerns.
I think batch-evals for non-sensitive data has great PMF here.
> So Apple won in some strange way again?
Heh, what did they win exactly? This is just a way for another company to extract value out of the single region of the world where Apple is a relevant vendor, and it happens to be the one where it's the easiest to pull people into schemes.
Yes. They never needed to participate in the AI race to zero.
Because they were already at the finish line with Apple Silicon.
> I don’t see a reason not to use this API. The only thing that comes to me is data privacy concerns.
The whole inference is end-to-end encrypted so none of the nodes can see the prompts or the messages.
Fun question: can some (part of it) be a crypto token that I can buy? :))
That would finally be a crypto thing which is backed by value I believe in.
Apple should build this, and start giving away free Macs subsidized by idle usage.
I thought this was Apple’s plan all along. How is this not already their thing?
I could imagine this working for the openclaw community if the price is right
That was my first thought too especially for talks that aren’t particularly important like daily digests of online things
I'm unable to download FLUX.2 models from `darkbloom models`
How does the inference work correctly if the payloads are encrypted?
Actually more useful than Bitcoin. Brilliant idea.
Why isn’t a MacBook Air M5 on the hardware list?
Because the model that generated that list was trained before the M5 came out.
no fans
Broken calculator or am I missing something here?
I mean, I'd be happy to buy a few used M2 Airs with minimal specs and start printing money but…latest (v0.3.8) tar doesn't contain image-bank or gRPCServerCLI dependencies so installer fails.
Like Fold@home but for profit!
I really want this to succeed
bittensor has something to say about this
Too much to read.
> That is not a technology problem. It is a marketplace problem.
I cringe every time I see this sentence structure. I know the joke is about emdashes, but the “Its not …. It’s ….” drives me crazy.
I see it everywhere now, it's even in videos and how people talk.
Maybe it's a Baader-Meinhoff phenomena and even a small difference in frequency feels overwhelming because you can't help but notice when it happens without noticing when it doesn't occur.
It learned it from humans....
In a little while, AI will reverse the order and everyone will be happy.
Tired: That is not a technology problem. It is a marketplace problem.
Wired: This is a marketplace problem, not a technology problem.
I think what irritates people is that there isn't a natural variation to whatever pattern it uses.
And I don't really know why both patterns don't appear equally.
Intuitively I just can't grasp how it can be so selective with the phrases it generates.
Only an em dash missing in between to be chefs-kiss-perfect.
It's not a sentence-structure problem. It's an effort problem.
But the real solution is to do this other thing. If you'd like I can give you the three-step guide to fixing the thing you asked me to fix.
/s
[dead]
[dead]
[dead]
[dead]
[dead]
[dead]