Can AI Keep Growing Exponentially? Let’s Ask SemiAnalysis — With Dylan Patel and Jordan Nanos

Anthropic

Across the shows: 12 speakers in other episodes, 8 bullish and 30 bearish calls, past 30 days

• Dylan Patel · SemiAnalysisBullish

  Anthropic is already profitable on compute, with about 75% inference gross margins, rivaling SaaS.

  Patel says Anthropic's revenue has grown about 10x year on year, while compute costs grew more slowly. Revenue now exceeds compute costs, training and inference combined. Inference gross margins are about 75%, close to Salesforce's, and the company has near-zero customer acquisition cost, unlike many SaaS firms. R&D spend has grown more slowly than revenue. Customers are sticky, since spend share barely moves when a rival model such as a Chinese one launches. He concludes the AI lab may be a better business than software.

  “We've seen Anthropic is now profitable in terms of compute cost minus revenue, or revenue minus compute cost, both training and inference combined.”at 13:20

• Dylan Patel · SemiAnalysisBullish

  Anthropic is a strong business, but a $10-20T valuation would need revenue multiples to collapse.

  Patel calls himself extremely bullish on Anthropic and says it could become the first $10T company. Still, Anthropic's revenue multiple has fallen from about 30x in private rounds to about 20x at IPO, and he thinks it is now nearer 5x. To justify $10T it would probably need around $1T of revenue, which he says could arrive in 2028 or 2029. At $20T, AI would have to be generating trillions of dollars for the whole economy. He frames the IPO as a bet that valuations compress as revenue catches up, with downside if it does not.

  “for them to be a 10 or $20,000,000,000,000 valuation company, especially 20... First of all, their multiples on revenue are gonna have to collapse quite a bit.”at 26:29

• Jordan Nanos · SemiAnalysisBullish

  Anthropic could reach a $20T valuation via non-coding products, though compute build bumps are likely.

  Nanos says the optimistic case is that Anthropic's revenue at $20T valuation levels would not come from per-token coding subscriptions. It would come from other industries such as drug development, similar to Eli Lilly's GLP-1 returns. He would trust Anthropic to bring a cancer drug to market. He still expects many bumps, since serving that demand and training models requires building compute, and the neoclouds are central to that build-out on the road from a $2T IPO to a $20T valuation.

  “I think that they get into other areas of business. And I think I would trust Anthropic quite a bit to bring, like I said earlier, a cancer drug to market”at 28:38

• Dylan Patel · SemiAnalysisNeutral

  Anthropic is holding back its strongest model, releasing only a weakened version, and may ship one before its IPO.

  Patel says Anthropic had its Mythos model ready in February. Its safety cards mention two other models released since then, but the full Mythos has not been released publicly. The public version is a neutered model called Fable, which took many months to appear. Glasswing, a version of Mythos, was limited to some sectors and neutered in some areas. Labs keep months of progress internal, because releasing a model lets competitors distill it. Patel expects Anthropic to release a new model in the coming weeks, ahead of its IPO.

  “Anthropic had [Mythos] ready in February... and publicly, we still don't have full [Mythos]. We have a neutered version of it called Fable.”at 59:11

OpenAI

Across the shows: 11 speakers in other episodes, 3 bullish and 14 bearish calls, past 30 days

• Dylan Patel · SemiAnalysisBullish

  OpenAI should reach compute-level profitability within two or three quarters.

  Patel says OpenAI gets to compute-level breakeven within the next couple of quarters, maybe two or three, based on what SemiAnalysis expects in its tokenomics model. He places OpenAI alongside Anthropic, which he says is already profitable on compute minus revenue, to argue the two AI labs are already at or near breakeven on their model economics.

  “And OpenAI gets there within the next couple of quarters, maybe two or three quarters, based on at least what we expect in the tokenomics model.”at 13:36

• Dylan Patel · SemiAnalysisNeutral

  OpenAI is keeping its newest models internal, such as Bell, and will likely release one soon.

  Patel says OpenAI finished its Astra model a couple of months before its release, and it took a while for Astra to come out. Its new model, Bell, was shown in a blog post with math benchmark results, but it has not been released. He says that makes it clear OpenAI is using it internally. Labs have already been holding back months of progress from public release, and the gap to Chinese open-source models is kept at about six months publicly. He expects OpenAI to have to release a new model in the coming weeks because of competition.

  “And their new model, Bell, they even showed a blog where they showed the performance of it in math... and they have not released it. So clearly they are just using it internally, not externally.”at 59:47

Alphabet

Across the shows: 5 speakers in other episodes, 3 bullish and 2 bearish calls, past 30 days

• Dylan Patel · SemiAnalysisBearish

  Google DeepMind's share of Google compute is shrinking, leaving it with the least compute of the top three labs.

  Patel says Google is producing much more compute each year, but the share going to DeepMind is falling. By the end of this year, DeepMind will have the least compute of DeepMind, Anthropic and OpenAI. Next year, OpenAI and Anthropic are each set to add five to ten gigawatts, while Google adds only a few gigawatts. That makes Google's compute deficit worse. He says talent is also moving to Anthropic, OpenAI and Meta, which compounds the problem. He notes Google has the talent to catch up, but more compute is what it would need.

  “It's actually falling as percentage, which is really interesting because now if I look at just DeepMind versus Anthropic and OpenAI, DeepMind has the least compute of the three by the end of this year.”at 55:57

• Jordan Nanos · SemiAnalysisBearish

  Gemini trails the top coding models, so DeepMind needs a niche to fund frontier research.

  Nanos says that after the recent Gemini release, Gemini is not competitive on coding with the top two or three models. Nobody wants the third, fourth or fifth best coding model, so Google needs to find a niche that drives Gemini token sales. Without token revenue, he argues, DeepMind will struggle to keep competing on frontier AI research. The argument is based on the gap in coding performance and the economics of token sales.

  “Like for Gemini, I mean, nobody wants the third or fourth or fifth best coding model. So they need to find a niche that drives Gemini token sales...”at 57:05

Meta Platforms

Across the shows: 8 speakers in other episodes, 12 bullish and 3 bearish calls, past 30 days

• Dylan Patel · SemiAnalysisBullish

  Meta is solvent and can stop buying new capacity; its cash flow covers signed commitments.

  Patel rejects the Goldman-cited claim that Meta has committed corporate suicide on debt. Meta can pause new spending at any time because much of its data center plan is not yet committed, and its cash flow can easily fund the contracts it has already signed. He calls it completely solvent, and says the real question is what revenue it earns on top of this spending.

  “Meta at any point can just stop buying new stuff and their cash flow can easily pay for all the stuff that they've signed.”at 11:46

Amazon.com

Across the shows: 3 speakers in other episodes, 0 bullish and 2 bearish calls, past 30 days

• Dylan Patel · SemiAnalysisBullish

  Amazon's AI infrastructure investments are all profitable today; Bedrock AI lifted gross margin.

  Patel argues that Amazon, the largest builder of AI infrastructure, earns a return on all of it today. Its customers, such as OpenAI and Anthropic, are solvent and carry strong gross margins. AI revenue from Bedrock raised Amazon's gross margin in its most recent quarter, which supports his view that the buildout is paying off on the supply side.

  “If you look at like Amazon, for example, the biggest builder of AI infrastructure, all of their AI infrastructure investments are profitable today.”at 13:42

NVIDIA

Across the shows: 12 speakers in other episodes, 11 bullish and 3 bearish calls, past 30 days

• Jordan Nanos · SemiAnalysisNeutral

  NVIDIA's neocloud backstops are a pragmatic supply-chain strategy, not a nefarious bubble-inflating scheme.

  Nanos says Semi Analysis tracks about $588B of off-balance-sheet backstops NVIDIA supports for neoclouds, against revenue approaching $500B, which he notes is about 2.5% of US M2. He argues NVIDIA wants a diverse customer base and does not want to depend on three or four hyperscalers. Because neoclouds need investment-grade off-takers to raise debt, NVIDIA signs on as that off-taker, then sells the GPUs to other buyers at higher prices when its own research need is smaller. He says NVIDIA's own internal compute is a very small share of revenue, and that the backstops also cover construction, memory suppliers, fabs and industrial equipment, since those are the long poles in neocloud projects. He defends the company against claims of a bubble-keeping motive, pointing to its research output, such as the most models published on Hugging Face and its software CI.

  “For everybody who says that NVIDIA is doing this for some nefarious purpose or to keep some bubble going, I'd encourage you to look at the quality of NVIDIA's research.”at 40:52

Oracle

Across the shows: 3 speakers in other episodes, 0 bullish and 2 bearish calls, past 30 days

• Dylan Patel · SemiAnalysisUncertain

  Oracle is a top neocloud on service quality, but execution delays and fixed-margin contracts weigh on the stock.

  Patel rates Oracle as one of the top neoclouds on ClusterMax quality of service for managed clusters, but says that is separate from execution risk. Semi Analysis told clients in April that Oracle's New Mexico data center for OpenAI would be delayed because pipelines were not being built. Oracle later declared force majeure and its stock fell. The delay could last multiple years. He also notes Oracle signed massive OpenAI deals last year at fixed margins, so it cannot capture higher GPU rental prices. Its debt raise spooked the market even though Oracle is solvent. He says these issues can hurt the stock badly, but he still thinks the company is fine.

  “But it is an objective fact that we think Oracle is one of the top neo clouds in terms of quality of service.”at 48:48

Nebius Group

• Dylan Patel · SemiAnalysisBullish

  Nebius's short contracts let it raise GPU rental prices, unlike peers locked into long deals.

  Patel says Nebius has the shortest average contract length among the top clouds in ClusterMax. That lets it reprice upward as GPU prices rise, and it has done so to take advantage of the market. He contrasts this with CoreWeave and Oracle, which have good cluster management, reliability and security but signed mainly long-term contracts. Those contracts left them unable to capture the same pricing upside.

  “they have the shortest average contract length of the top clouds in ClusterMax. So Nebius has been able to turn around and really jack up pricing and take advantage of this.”at 51:32

CoreWeave

Across the shows: 1 speaker in other episodes, 0 bullish and 1 bearish calls, past 30 days

• Dylan Patel · SemiAnalysisBearish

  CoreWeave's long-term contracts keep it from capturing higher GPU rental prices despite good operations.

  Patel says CoreWeave has good cluster management, reliability and security, but it locked itself into mainly long-term contracts. As a result, it was not able to take as much advantage of rising GPU rental prices as Nebius, which has shorter contracts. The point rests on the contract structure, and it is a relative comparison rather than a claim about absolute financial distress.

  “CoreWeave and Oracle, despite having also good cluster management and reliability and security and all this stuff, were not able to take as much advantage because they locked themselves into long term contracts primarily.”at 51:45

AI infrastructure & data centers

• Dylan Patel · SemiAnalysis

  US AI capex is far larger than commonly estimated, around 5-6% of GDP, serving global demand.

  Patel says the widely shared chart understates the buildout. US capex next year is about $2T including the whole supply chain, and with about $300B+ of GDP-related spending the total is roughly 5-6% of GDP, not 2-3%. Much of this US capex serves foreign users, since most Europe and Asia inference runs in American data centers and all training happens there, so the economic footprint is larger than in railroads or fiber, where capex stayed in the country where it was deployed.

  “CapEx across The US will be on the order of $2,000,000,000,000, not just data centers and chips, but also all the rest of the supply chain that people are investing in.”at 3:18

• Jordan Nanos · SemiAnalysis

  Capex growth cannot keep accelerating; neoclouds struggle to organize and finance more gigawatts.

  Nanos argues that triple-digit capex growth cannot accelerate further because the supply chain runs out of room. The biggest neoclouds and hyperscalers find it hard to organize and marshal the resources to build another gigawatt on top of current projects. They also find it hard to raise capital, even though returns on these investments are strong, because the required sums are measured in trillions of dollars.

  “you can't just have triple digit growth and accelerate the second derivative even further... You run out of things.”at 7:51

• Jordan Nanos · SemiAnalysis

  Neocloud contracts run five to six years without monthly cancellation; pause risk is in self-build sites.

  Nanos argues the neocloud contract risk is overstated. Hyperscalers and frontier labs sign five- to six-year terms at a fixed hourly rate, taking full capacity, with no arbitrary monthly exit. Neoclouds secure land, power and sometimes chips before signing. The pause risk sits with hyperscalers' own self-built campuses, which can be left as shells. Construction is a small share of a new project's cost, while GPUs, HBM and optics dominate the bill, so the spending that could be cut is less firm than the headline numbers suggest.

  “Generally speaking, they're signing five year, six year term lengths and they're off taking the full capacity.”at 22:28

• Jordan Nanos · SemiAnalysis

  Modeling shows AI compute demand still outstrips supply, with chip and data center build limits binding.

  Nanos says all of SemiAnalysis's modeling shows demand outstrips supply. Hard constraints on how many chips can be produced, how many data centers can be built and how fast capacity comes online mean neoclouds and hyperscalers cannot catch up to demand. He cannot say whether demand is 10%, 50% or infinite above supply, but the gap has not closed. Any slowdown would require demand to weaken first.

  “all of the modeling we can do basically shows that demand outstrips supply.”at 24:51

• Dylan Patel · SemiAnalysis

  Most neoclouds have weak security; open-source Chinese models can hack many of them.

  Patel says Semi Analysis published a report titled that most neoclouds are bad at security. He says open-source models, especially GLM 5.3, can be used to hack dozens of neoclouds or get access to data inside them. He describes an incident where Semi Analysis accidentally saw other clients' storage and Slurm setups on a shared cluster, and says a dozen other customers on that cluster could have been training models. He says this could expose what they were training, and that a non-American government security agency was on a top-10-GDP country's cluster. He sees this as a real risk given that these clusters are used for training and running models.

  “Look, most of these clouds are terrible at security. You can get open source models to just hack Neo Clouds or get access to stuff within Neo Clouds.”at 44:05

• Jordan Nanos · SemiAnalysis

  Sovereign governments treat neocloud compute like telecom infrastructure they want to control.

  Nanos says people use the term sovereign AI when raising money because they see compute as infrastructure. He compares it to the telecommunications build-out, where nations treat the infrastructure as an asset to control. These countries want to be net token importers or exporters, which depends on data centers filled with GPUs. He argues it is therefore no surprise that government research and intelligence agencies share clusters with other users. A country with only one facility holding GB300 NVL72 systems is a seller's market, so governments end up on shared infrastructure.

  “This is just like the telecommunications build out, where this infrastructure is viewed by the sovereign nations that are deploying it and are involved in this as an asset that the country would like to control and have sovereignty over.”at 45:40

• Jordan Nanos · SemiAnalysis

  Neocloud service quality lags because the seller's market rewards fundraising and speed over reliability.

  Nanos says Semi Analysis tracks 323 providers and only a small share are large enough to list and pass its bar. The bar is low, since basic failures on security and reliability are dead simple to fix. Many providers have not improved in the nearly two years since the first ClusterMax. He says a seller's market means companies prioritise raising money and deploying chips quickly over the quality of the underlying service that customers want. That lets businesses grow without delivering a service customers enjoy and recommend.

  “the dynamic of it being a seller's market means that the priority for these companies is how can they raise as much money as they can, how can they build data centers as fast as they can...”at 53:20

AI: overall outlook

• Dylan Patel · SemiAnalysis

  Consensus AI revenue requirements are too low; OpenAI and Anthropic alone could reach $1.5T by 2027.

  Patel argues that revenue estimates for the buildout (Bain's $6T by 2031, the $3.5T by 2032 figure) are too low. Infrastructure assets are depreciated over about six years for GPUs and about fifteen for data centers, so a $1T capex wave needs roughly $1.5T of cumulative revenue, about $250B a year. He thinks OpenAI and Anthropic can reach about $600B+ combined by December next year and around $1.5T by end of 2027. Revenue lags capex, and current AI revenue reflects infrastructure built in prior years, so the payback looks achievable on that timeline.

  “I think everyone's numbers for revenue requirements are quite low given the case where the build out is gonna go.”at 10:41

• Jordan Nanos · SemiAnalysis

  Frontier labs could earn AI returns far beyond coding, in drug discovery, robotics and self-driving.

  Nanos argues that AI revenue should be measured more broadly than today's coding and per-token pricing. If frontier models can do most knowledge work, labs could enter robotics, self-driving, drug discovery and life sciences, earning on products such as a cancer drug or a self-driving vehicle. He says people lack the imagination to see returns beyond coding subscriptions and token pricing, and only frontier labs can compete there.

  “you can imagine them going into a whole bunch of different industries beyond coding and having returns on their investments similar to what they've seen in coding.”at 14:32

Frontier AI labs & model race

• Dylan Patel · SemiAnalysis

  Frontier labs can do little to slow pace; the race with rivals and China forces continued speed.

  Patel argues the labs cannot slow down much. Anthropic and OpenAI each refuse to slow because the other will not, and China is not far behind. Pacing the frontier would be responsible only if the risks were existential, and he doubts the labs will pace themselves much. He says any slowdown would most likely come only after major security incidents.

  “I think like there's very little pacing they can do given the race that they're in.”at 33:40

• Jordan Nanos · SemiAnalysis

  Labs will keep releasing minimum viable models to stay ahead of Chinese open source, pacing themselves.

  Nanos says the incentives are unclear: API per-token margins favour releasing more models, while keeping models internal favours end-user applications. He thinks Anthropic is only slowing itself down, since nobody else is pacing the frontier. He bets that OpenAI and Anthropic will keep releasing minimum viable models to stay ahead of Chinese and other open-source models. He also says the best incentive would be using AI to solve clearly valuable problems like cancer, which could be profitable too.

  “my bet is that we continue to see the minimum viable models released by both OpenAI and Anthropic to keep a lead over Chinese and the rest of the open source ecosystem.”at 1:02:01

AI & jobs

• Dylan Patel · SemiAnalysis

  AI will accelerate the falling labor share of income as value accrues to chips and data centers.

  Patel says labor's share of GDP has fallen steadily since the 1970s, with value shifting to capital. AI tokens are essentially capital embodied in chips and data centers that produce labor, so AI should supercharge that trend and make labor's share fall faster. He still hopes labor dollars won't shrink, because the wider economy grows fast enough. He cites chip design, where US designer headcount has been flat for two decades while the value created exploded, as an example of AI letting firms grow revenue and margins with the same number of people.

  “What AI does is it's gonna supercharge that trend... Long term trend of labor share versus capital share because AI tokens are general... Basically just capital”at 17:23