Every relevant Silicon Valley CEO (Jensen, Elon, Sam, Satya, Dell, etc, etc, with the only exception being Dario) Fund Manager/VC (Gerstner, Baker, Andreesen, Gurley, Sacks, Chamath, etc, etc) is full throated endorsing open source AI as a key strategic technology for America that must be fostered/cultivated rather than cut off by regulation. How does an investor express a trade that benefits from this? The obvious one is Bittensor. Bittensor is an open, permissionless, decentralized ecosystem designed to fund and coordinate open source compute protocols, pre/post-training, AI apps/businesses, AI research, etc (think: open source AI startups) Bittensor’s TAO token is a liquid index designed to capture the value these open source AI start ups are creating. It’s a hard money asset with the exact same tokenomics as Bitcoin. It currently trades at 2B mcap. This is the trade of the decade sitting in plain sight.

Interesting

What is the TAM for inference providers that serve frontier-grade intelligence (Kimi K3.1 Pro) at the lowest possible cost in a private fashion?

The future of enterprise AI is local / decentralized

Probably nothing

Great explainer on Gamma Tokens from @markjeffrey Bittensor mines for intelligence and then tokenizes it. Tokenized Intelligence.

Started writing about the upgrades I felt would be needed in order for PRL to have any long term durability and this ended up getting quite long. Fair warning: long form writing is always quite helpful for me to really flesh out/test my own ideas so feel free to scroll past this (except for you, @tulipking). Before getting to what changes are necessary I feel it’s important to step back because they are only going to make sense once I’ve explained what they are meant to fix. I also want to preface this by stating for the record that PRL has the potential to be a good token, it didn’t have to be this way, but the “TAO bad = PRL good” takes on the timeline pushed me to write it down. Also none of what I am suggesting here is that the engineering behind the chain is a scam. For starters, I agree that Pearl attempts to tackle a very real problem for Bitcoin (the security budget). Bitcoin pays its miners close to a billion dollars a month, but the subsidy keeps halving (next in 2028), so barring a massive repricing of Bitcoin (possible), the miner reward is shrinking at the same time there’s a viable alternative for the datacenters (inference/training). This year hashprice hit record lows, difficulty took one of its biggest drops ever in June, and there’s been reports that some miners could be getting up to 70% of their revenue from AI by year end. Hashrate is back near 1 ZH/s this month, so it isn’t a crisis, but the trend is not our friend here. I do think Pearl’s answer to this problem is clever. Instead of competing with AI for the same power and datacenters, harness that same AI work by using the matmuls a datacenter already runs to serve inference to secure the chain. Keep everything else about BTC the same and voila you have, PRL, the “AI Bitcoin.” A new digital store of value that solves the security budget problem. But, actually, you don’t have a new Bitcoin just because you made this maneuver. There are multiple reasons for this which together I dub the “Bitcoin immaculate conception” below: (1) A 1/1 origin story. Bitcoin launched in the middle of the GFC, with a Times headline about a second bank bailout embedded in its genesis block. It had a villain and a purpose, and nobody gets to tell that story twice. (2) A founder who vanished. Satoshi was anonymous, went quiet in 2011, and left no company, no treasury, and nobody steering, so there’s no one to pressure, subpoena, or wait on. Satoshi’s ~1.1M BTC (about 5% of all bitcoin) has sat untouched for over fifteen years, so the market treats it as burned rather than as supply waiting to be sold. (3) Free mining. Early miners were hobbyists running CPUs (then gaming GPUs) on computers they already owned, so there were no real costs and no reason to sell. Real ASICs didn’t come online until ~2013. (4) Nothing to sell into. The coins had no price for most of the first year, and half of all bitcoin was mined before the whole network was worth $1B (which didn’t happen until 2013). The front loaded supply spread out cheaply to people who wanted it instead of landing on miners who needed to dump it. (5) Nothing else to buy. For about two years Bitcoin was the only cryptocurrency (Litecoin came in 2011), so anyone who wanted in had exactly one option, and it was already the default by the time alternatives showed up. (6) It became the Schelling point. Everyone agrees it’s the digital money, and that agreement is the asset. The meme now compounds through ETFs, corporate treasuries, and even a US strategic reserve, and every new class of holder makes it harder to dislodge. (7) It’s Lindy. It survived Mt. Gox, three 80%+ drawdowns, China banning mining, and a civil war over block size. Every year it doesn’t die makes the next one more believable. Said differently, it’s a combination of a 1/1 founding story, unique distribution dynamics, and a heavy dose of memetics. PRL has almost none of this. Its founders are doxxed, their miners are predominantly professional miners with substantial OpEx, markets they can sell into fairly easily, and on the other side there are many other coins people can buy. Honestly the list of real stores of value in crypto is Bitcoin at best. ZEC is probably making the best push (it’s basically Bitcoin with privacy), but it isn’t there yet. [As a brief aside: I’ve seen many PRL bulls critique Bittensor for its rapid updates/refinements to the protocol, touting ossification / simplicity as an essential element to being a digital SoV as if updates/change are something PRL can avoid. I’m not sympathetic to this argument so I did not include it above. Pearl has to keep upgrading as well, because its work only stays useful if it keeps up with whatever AI actually runs. Their own roadmap post calls Pearl a new kind of blockchain “evolving alongside AI.” Matmul itself is stable, but the number formats AI runs it in change constantly. Every Nvidia data center generation since 2017 has added a lower precision format (FP16 with Volta, BF16 and TF32 with Ampere, FP8 with Hopper, FP4 with Blackwell). Pearl put AI in the consensus layer, so every one of those changes is a fork. It’s already done four since April, including turning MoE proofs on in June and back off in July after miners used them to fake hashrate, which matters since most frontier open models are MoE. The FP8 draft also has the verifier recompute work bit for bit for specific GPUs, which is why its reference miner drops RTX 30s, 40s, and A100s creating a changing hardware whitelist. Pearl’s verifier needs a new entry every chip generation. Bitcoin could mostly ossify because its work is useless. Double SHA-256 in 2009 is double SHA-256 today, because there’s nothing outside the chain it has to keep up with. Pearl made the work useful, and useful is a moving target. If it ossifies, the work drifts away from what AI actually runs and you’re left with GPU proof of work. If it keeps up, it forks forever. Plus even BTC will have to eventually deal with quantum, cryptographic breakthroughs, and potentially the privacy angle. Bottom line - ossificaition is a meme.] Even Pearl’s own whitepaper doesn’t claim it can be a store of value on its own. It says that as a store of value, PRL is only worth something if people expect it to be adopted as a means of payment, and that without that assumption the token’s value and the mining “should clearly be zero.” So PRL needs what every coin since Bitcoin that’s pulled in durable outside capital has had, a fundamental reason to hold it. (1) ETH gave you smart contracts, and to use them you had to own ETH. (2) ZEC gives you privacy, basically a personal Swiss bank account. (3) HYPE generates real revenue and uses it to buy back the token, so the token captures the growth of the business (which is the same path the subnets on TAO are taking). PRL has none of these right now. There is no reason for durable outside capital to hold this token. Nobody is going to hold PRL to store their wealth. Because it lacks programmability, there’s nothing I can do with it, and the network generates no revenue (and even if it did, there is no value capture mechanism). The only people holding it today are speculators. The solution to this problem should be the easy part for Pearl. The whole point of proof of useful work is that useful work is happening, which means value is being created, so the token subsidizing that value should be capturing it. PRL doesn’t, because that value never touches the protocol. For example, Together’s customers pay Together in dollars, Together also mines PRL on the side, and the only thing connecting the two activities is that the same GPUs were involved. Every inference provider on earth could “duplex mine” and PRL’s top line wouldn’t change. More on this later. So as it stands today PRL as a network has barely any supply of, or demand for, useful work. But what it does have is sellers. The people the protocol pays every day are sellers by design, leveraging the token as a subsidy by selling it for USD to help improve the margins of their inference business. The usual reply I’ve gotten when I point this obvious dynamic out is that duplex miners will hold since the inference revenues they generate outside of the Pearl protocol covers their costs. I think that’s cope. The network advertises itself to these providers by touting a free subsidy in exchange for providing their MatMuls to secure the chain. My understanding is that there’s a ~5% cost to this as-is. The main partner, Together, has committed its PRL to a 25% discount, which means selling. Besides, the duplex guys aren’t the marginal hashrate right now anyway. Hashrate Index described most current useful work as inference nobody bought, and the easy way to mine is SRBMiner on a rented GPU serving zero inference. Those miners pay rent and power in dollars and sell to cover expenses. All of this mining emission is a lot to absorb when there’s nobody pays the protocol for anything, and it mints roughly a million PRL a day, about 305M more over the next year on ~348M circulating, so supply nearly doubles. At current prices that’s $1.7 to 2M of new PRL every day that somebody has to buy just to keep the price flat. Which, finally, brings me to your question on what iterations I think Pearl needs, because the fabled “marketplace” is the one everything else hangs on. It’s the only proposed mechanism by which a customer ever transacts in PRL, and, from what I can tell, there isn’t really a plan for it. The whitepaper admits tokens can’t be turned back into compute, says a marketplace is what would tie the token’s value to the value of useful compute (i.e. today nothing does), and treats on chain settlement of compute contracts as a future idea. The team’s own roadmap puts the marketplace and on chain settlement in the longer term (behind MoE and floating point support), there’s no launch date, and their open problems page calls using PRL to trade compute a concept that’s “far from fleshed out.” The changes it needs are big ones. The whitepaper version, where the chain settles compute contracts, needs programmability Pearl doesn’t have. So either they add it, which is exactly the kind of complexity you’re crediting them for avoiding, or Pearl Research Labs runs the marketplace as a company, at which point it’s a company selling inference with a coin attached (sound somewhat familiar?). Then there are the structural business problems that lie beyond the engineering. The primary issue is that Pearl doesn’t actually have anything to sell. The compute belongs to the miners, and they already sell it elsewhere. Together sells inference through enterprise contracts and OpenRouter, gets paid in dollars, keeps the revenue, and mines PRL on the side. A marketplace means putting Pearl between them and their customers and I haven’t seen a good answer for why they’d share those economics with Pearl. I’ve seen the argument that econmics don’t have to be shared citing the Pearl whitepaper which suggests that the PRL token will become the unit of account in this marketplace which will drive value to the token. But even if Pearl managed to get this marketplace up and running, settling inference purchases in PRL does not get anyone holding PRL. A buyer grabs PRL to pay for a job, the provider has dollar costs and sells on receipt, and the token gets held for minutes. This is a tale as old as time in crypto which @KyleSamani wrote up back in 2017 (the token velocity problem), and one of his fixes was a profit share (i.e. a revenue funded buyback, which is exactly what Pearl doesn’t have). Not to mention how difficult it would be to convince customers to buy PRL first before paying for inference. This is impractical and unlikely to scale as opposed to using USD. Which brings us back to sharing in the economics. If the marketplace settles in dollars instead, the best Pearl gets is a fee. Taking OpenRouter, with its dominant inference marketplace, as a best case scenario Pearl might be able to add a 5% spread on transactions routed through its marketplace and burns every dollar of it. To offset $500 to 600M of new PRL emissions over the next year you’d need roughly $10 to 12B running through this marketplace. OpenRouter is doing around $3B a year of spend (my back of the envelope from Sacra’s $160M revenue estimate). So a marketplace that doesn’t exist yet would need to be 3 to 4x OpenRouter just to hold PRL flat. Given the network effects OpenRouter has already built it will be a stark challenge bootstrapping the demand side for a marketplace to compete at that size/scale. [Note that a subnet doesn’t have this problem because it doesn’t need to own distribution. It can sell through OpenRouter and still capture the value, since the subnet is the business rather than the middleman. The miner’s output is the product, the subnet sells it, and the miner gets paid in the token the business runs on. A token captures value from a service in basically three ways: settlement (which velocity kills, see above), revenue funded buybacks or burns, or making people lock, stake, or burn it to participate. Pearl has none of the three live. Bittensor has the last two. Registering a miner costs TAO on most subnets, registering a subnet seeds its pool with TAO, and since alpha is priced in TAO, buying alpha means putting TAO into that subnet’s pool. Lium is the cleanest example of the buyback side. It rents out GPUs for dollars like anyone else, then uses the revenue to buy and burn its own token, about $1M in a single burn last month and close to $2.7M over the last six months. Their latest update puts them at ~$20mm ARR right now. Customers can pay in dollars all they want, because the subnet converts the revenue into its own token. Subnets don’t skim a fee off someone else’s business, they are the business, and the alpha token (i.e. the miner reward) is how they scale the business which is why the buybacks happen at all. In fact subnets like Chutes/Engy/Actual would actually be quite synergistic with a Pearl marketplace as they could sell inference there/mine PRL.] So, to recap: I don’t believe a marketplace that settles in PRL solves the issue. I also think that a marketplace where Pearl runs an OpenRouter style fee arbtirage is exceedingly unlikely to work. The real fix is to sell the inference the matmul mining is attached to, so the work that earns the block reward is the same work a customer paid for. Selling the inference requires verifying it. To do this will require substantial updating of the protocol. Right now the block reward can’t tell a paid job from a synthetic one. The usual argument is that this is OK and a miner serving real inference beats a synthetic miner in the long run, because the matmul gets done once and pays twice. In other words, because the inference provider is paid for the inferenace AND the PRL rewards they are less likely to sell their PRL than the synthetic miner who must sell to cover their OpEx. I think this argument is flimsy because it ignores the extremely likely probability that a legit inference provider witll themself mine synthetic matmul. Their inference revenue is the same either way, so it drops out of the decision. Take a hypothetical inference provider who also mines PRL: If his GPUs aren’t fully booked, synthetic is the only way to mine, and when they are booked, mining on the real thing slows down the product his customers are paying for (that same 4 to 5% from Pearl’s benchmarks, and Castle Labs puts today’s INT8 path at 10 to 15%). So even a completely legit provider is incentivized to keep his customer pipeline clean and mine synthetic in the gaps. And even at zero overhead (I read they’re targeting 0 to 3%), useful hashes are capped by how much paid inference exists, while synthetic hashes are capped by nothing except GPU supply. Pearl’s whitepaper admits the network attracts compute that isn’t doing useful work, and a June preprint found no identifiable inference code in the dominant mining software on the pool it measured (Link: arxiv.org/html/2606.04819v1). The point is, if there is to be a legitimate inference business built around Pearl, proof of verified inference will be absolutely essential. There are two ways to verify it, and both would obviously require loads of upgrading. (1) One is proving it on chain (they’re working on a L2 for this with Attestable apparently, but I haven’t seen a test date), and it has to go further than people assume. Pinning the weights to a real model isn’t enough, since a miner can multiply real weights by junk activations, which is just running a model on prompts nobody sent. You’d need proof the inputs came from a real request and that someone actually paid for it, otherwise miners pay themselves for jobs made of synthetic matrices and the protocol ends up charging a minimum burn per job (Bittensor’s registration burn with extra steps). (2) The other is validators judging which work is real, which is the route Bittensor took. It runs the AI off chain and uses a consensus mechanism that prices work by distilling the scores of independent staked validators, rather than trying to prove usefulness cryptographically. Validators bring their own problems (collusion, weight copying), and that’s a fair criticism of Bittensor. But Bittensor contains the failure. When a subnet gets gamed, that subnet loses emissions and the chain keeps running. When Pearl’s on-chain verification got gamed with MoE, the fix was a chain wide fork. And all of this stacks on top of the format forks I mentioned above. Whichever route the team selects for, Pearl would have finally achieved a state of being that would give durable outside capital a reason to hold. The chain would be simultaneously selling inference and securing itself with those matmuls while driving that value created back to the PRL token. This would be quite interesting for me personally. Like anything else, there are tradeoffs. The tradeoff of course in the scenario detailed above would be that hashrate would plummet. Today’s hashrate is inflated by rented GPUs multiplying random matrices. Future projections rely on the assumption that more and more large neoclouds will come on and duplex mine PRL. If the block reward only paid for inference somebody bought from the protocol (either OpenRouter style or Subnet style), security shrinks to the size of paid demand, which today is basically the Together endpoint on a Pearl modified Gemma plus Pearl’s own discounted credits. There’s no guarantee they ever sell enough inference to be the high hashrate chain they are now. That’s fine, it just means far fewer people mine it, but it doesn’t square with the pitch of out-hashing Bitcoin. I can imagine that this would seem incredibly unattractive to PRL bulls. But if they stay on the current path, they get a lot of hash and not much value. Hashrate ultimately follows price, not usefulness. If nothing holds the price up, miners leave, security drops, and the reason to hold drops with it. Even the hash that’s there is contingent. Together serves inference whether Pearl exists or not, so if PRL falls far enough it switches mining off and its business barely notices, while a Bitcoin miner’s whole business dies without the chain. I’m not saying that spiral is guaranteed, but nothing in the design stops it. Which is ironic, because this is the exact problem Pearl was supposed to solve. Its answer to who pays for proof of work was the AI industry, but the AI industry pays Together, not Pearl. Without value capture the answer is the same as Bitcoin’s (whoever buys the token), minus Bitcoin’s monetary premium to lean on (this is given not earned). Every fix above is years away from an engineering standpoint (if it happens at all). In the meantime there is a major glut if supply already mined / coming online that will have to be distributed before any durable outside capital can be attracted into the network. PRL may have been a “fair launch”, but in practice the emissions curve was so front loaded that more than half of all PRL in existence was mined in the first couple of weeks before there was even a market price for the token. My understanding is that a third of all circulating supply was mined in the first six days at near 0 difficulty. In week one, the people who knew about Pearl were people close to the project pretty much by definition, and Pearl’s own whitepaper says the low starting difficulty was meant to give early miners a fair shot and to give the core team resources for the R&D needed to keep up with new AI architectures and hardware (which is the forever upgrade problem I mentioned earlier). I had Opus 5.5 sample launch week off Pearl’s explorer (400 blocks, every 100th of the first 40,000) and grouped addresses that were later spent together, which means one wallet controls them. One operator with at least 118 addresses won about 61% of blocks in the first ten hours and took roughly 33M PRL over launch week, around 9.5% of today’s supply. Its known addresses still hold about 98% of it, roughly 30M PRL, which is about eight and a half days of PRL’s entire reported trading volume. Across the launch addresses I could trace, at least 57M PRL (16% of supply) has never moved. No vesting, no disclosure, and since this is a sample those numbers are floors. I can’t tell you who the operator is (an outside operator with early access would look identical on chain), but who it is matters less than the fact that it hasn’t sold. Today’s miners sell, and the launch cohort mostly hasn’t, which is worse in a way, because it’s all still sitting there. I have no problems with this design of emissions curve, but it does front load massive supply and exasperate the overhang problem. Real revenue can absorb an overhang (that’s the whole point of HYPE’s buybacks), but that’s exactly the revenue Pearl doesn’t have. So even if you believe the whole roadmap, the rational move is to wait until issuance slows and the early bags get distributed, because a “store of value” where you’re the exit liquidity isn’t protecting anybody’s wealth. To recap and tie it back to the original question (what iterations do I think are needed?): Pearl needs a marketplace with no real plan behind it yet, verified inference that goes all the way down to who paid for the request, a buy+burn program. Not to mention the forks every time the hardware/industry standard moves. Pull all of that off and what you’ve built is basically a subnet running on its own L1, years behind the subnets already doing this with paying customers. I’ll grant it would be a subnet with one thing no subnet has, which is inference that actually secures a store of value. This would be immensely valuable. I just don’t see the path from here to there yet. If one shows up I’ll happily be a buyer, ideally after the early miners distribute their bags. My best guess is this will be a next cycle play.

Really interesting convo below from this weeks all in on why a16z stays away from the neoclouds, and it made something click for me about Bittensor. @bhorowitz point is that scaling compute on debt is a dangerous arbitrage business, and he lived it in 99 when bandwidth was $1,500/megabit and the only way to get it was a 3 year contract. By 2001 it was $50 - If you were holding that on borrowed money you were dead. His case for $CRWV being different is that AI compute is in a massive shortage, so GPU and data center prices keep rising. Maybe he's right. But the neocloud model is still a levered bet that GPU scarcity outlasts the debt, because they borrow against the GPUs and the lenders get paid on a schedule whether or not the market cooperates. A compute subnet on Bittensor is basically a decentralized neocloud that doesn't need any of that. There's no GPU purchase to finance (miners show up with their own hardware) and no 3 year capacity contract to get stuck holding if prices move. The token + the network do the job the debt does for CoreWeave. Emissions bootstrap the supply side, and as demand and revenue grow the subnet puts that value back into its token, which raises what miners earn and pulls more compute onto the network. The token is the lever that scales how much compute you can harvest and coordinate. So if GPU pricing ever pulls a Loudcloud, the subnet doesn't blow up, its incentives just buy more compute than they used to. With no debt there's nothing to default on, and nobody can call a loan or force a fire sale. The risk you take on instead is token volatility, which you manage with a treasury (keep cash on hand and use it to support your token through drawdowns). A bad market can starve a subnet for a while, but one with real demand underneath it can ride that out. The tradeoff is a slower, more gradual build. A neocloud can borrow its way to a giant cluster fast, while a subnet's capacity grows with the size of the network and the token. What you get in exchange is a far more durable business, mostly because of where the supply comes from. You're tapping a bunch of long tails of compute that a neocloud can't really get at: >people who own their own hardware and want to monetize it when they're not using it >slack capacity in data centers (nobody runs at 100%), which can get sold into the subnet's spot market instead of sitting idle >sunk cost hardware, i.e. paid off GPUs that can't get a contract for anything else >operators in places with cheap power that no lender would ever underwrite a data center for Once a card is paid off the main ongoing cost is power, which even on an H100 is something like 5-10 cents an hour depending on where you are. So a lot of this supply is pretty price insensitive. They make more when the token is higher, but a 50% drawdown in miner rewards eats into their margin and plenty of them keep mining anyway. That gives you a structural supply floor that doesn't really care what the token (and therefore the miner reward) is doing day to day. Then as network value and the token go up, you progressively unlock compute that's more price sensitive (financed GPUs that still have to pay themselves off, US data centers with pricier power, etc). Each one joins once mining becomes economic for it, and you get a flywheel: token up → more of the world's compute profitable to plug in → more capacity to sell → more revenue → more value back into the token → repeat. Every GPU that joins also makes the network cheaper and more reliable to buy from, so you get a real network effect on top of removing the debt risk. And in the limit it can reach more compute than any single hyperscaler could ever build or finance, because the supply side is permissionless and global. You start by tapping into the world's edge compute, and over time the rest of the world's compute comes online as it makes economic sense for whoever owns it.