Anthropic overtook OpenAI in annualized revenue this spring, hitting a $30 billion run rate against OpenAI’s roughly $24 billion — and it did so while planning to spend about a quarter as much on model training. That combination is the most important signal in AI right now, and it points somewhere most coverage missed. The verdict is this: the winning AI business model is capital-efficient enterprise inference, not consumer-subsidized frontier scaling — and that shift is the strongest structural argument yet for decentralized compute markets, because the industry’s binding constraint is becoming cheap, verifiable inference capacity rather than the next $100 billion training cluster.
Read that carefully, because it inverts the dominant narrative. For three years the AI story has been about who can raise the most capital to build the biggest training run. Anthropic just demonstrated that the company generating more revenue is the one spending dramatically less on exactly that. If capital efficiency is winning, the entire thesis for centralized, hyperscaler-owned compute weakens — and the case for open, market-priced compute strengthens.
The numbers that flipped the script
The crossover is real and recent. In April 2026, Anthropic reached a $30 billion annualized run rate, up from $1 billion roughly fifteen months earlier, while OpenAI’s own figure sat near $24 billion, about $2 billion per month. Epoch AI had projected the crossover for around August 2026; it arrived early. Anthropic has grown roughly 10x per year since crossing $1 billion, against OpenAI’s 3.4x.
The revenue mix explains why this is durable rather than a quarterly blip. Anthropic draws roughly 85% of revenue from enterprise and developer customers — more than 500 companies now spend over $1 million a year, and eight of the Fortune 10 are customers. OpenAI’s mix is the mirror image: heavily weighted to ChatGPT consumer subscriptions, where the overwhelming majority of users pay nothing. One company sells a high-margin input to businesses that turn it into value; the other subsidizes a mass consumer product and hopes to convert it.
Then the cost side, which is where the thesis lives. OpenAI’s compute spending is projected to reach $121 billion in 2028 alone, with the company burning roughly $17 billion in cash annually and not expecting positive free cash flow until 2029. Anthropic’s training costs are projected to peak around $30 billion in 2028 — roughly 4x less — with profitability targeted for 2028 or 2029. More revenue, a quarter of the training spend. That is not a rounding difference. It is two opposing bets on what AI economics reward.
Why capital efficiency, not scale, is the winning bet
The last three years trained the market to believe that the biggest training run wins. Anthropic’s results complicate that. It is generating more revenue with far less training capital, which means the marginal dollar of value in AI is shifting from training frontier models to serving them profitably at scale. Enterprises do not pay for the size of your last training run; they pay for reliable, affordable inference wired into their workflows.
This matters because training and inference have opposite cost structures. Training is a lumpy, centralized, capital-destroying event — one enormous cluster running for months. Inference is a continuous, distributable, capital-returning operation — millions of small requests that can, in principle, run anywhere there is a GPU and a network connection. As the industry’s revenue tilts toward inference, its cost base wants to tilt toward whatever supplies inference capacity most cheaply. Centralized hyperscalers are not obviously the cheapest supplier of that; they are the most convenient one, which is a different thing.
OpenAI’s own financing behavior underlines the strain. A company spending $121 billion on compute in a single year and losing $14 billion in 2026 is, functionally, a compute-financing vehicle wrapped around a consumer app. We argued this directly when we broke down how OpenAI’s $122 billion round was really a compute-financing deal. Anthropic just showed there is another way to run the race — and the cheaper way is currently ahead on revenue.
The constraint is moving from training clusters to inference supply
Follow the bottleneck. When the scarce resource was the ability to assemble a giant training cluster, capital and hyperscaler relationships were the moat, and that favored whoever could raise and spend the most. But if capital-efficient inference is what actually converts to revenue, the scarce resource becomes affordable, verifiable, geographically distributed compute for serving models — and that is a market, not a single cluster.
Two forces push in the same direction. First, the ongoing memory and DRAM shortage has made high-end centralized capacity more expensive and harder to secure, raising the price of the convenient option. Second, inference workloads are far more parallelizable and latency-tolerant than training, which makes them a natural fit for distributed networks that would be hopeless for a synchronized training run. The workload that is growing is precisely the one that decentralizes well.
None of this says training stops mattering or that hyperscalers vanish. It says the growth of the market is moving toward the layer where an open, market-priced compute supply can actually compete on cost — and Anthropic’s efficiency lead is the clearest evidence that cost, not raw scale, is what the revenue rewards.
The Web3 angle: decentralized compute has its demand case now
Decentralized compute has spent years searching for a demand story stronger than ideology. Anthropic-versus-OpenAI supplies one: if the profitable model is capital-efficient inference, then networks that undercut hyperscaler inference pricing have a real buyer. The relevant projects are specific.
Render Network (RNDR) built a marketplace for GPU rendering and has extended toward AI inference workloads, matching idle high-end GPUs to paying demand at prices set by an open market rather than a cloud rate card. Akash Network (AKT) runs a decentralized compute marketplace where GPU capacity is bid for directly, routinely undercutting centralized cloud pricing for comparable hardware. io.net (IO) aggregates GPUs into clusters aimed specifically at machine-learning inference and training, targeting exactly the cost gap this shift creates.
Further out on the risk curve, Bittensor (TAO) is building an incentive network for machine intelligence itself — paying participants in a token for producing useful model outputs, an attempt to decentralize not just the hardware but the model-serving layer. Whether TAO’s specific mechanism holds up is an open question, but the direction matches the thesis: value accruing to distributed inference rather than centralized training.
The bridge to the token investor is the one we have made before in tracking how the AI compute trade is rotating: as inference demand grows and centralized capacity stays expensive, entities with power, cooling, and GPUs — including repurposed bitcoin miners and decentralized GPU networks — become the marginal suppliers. Anthropic’s win is not a crypto story on its surface. Underneath, it is the clearest demand-side argument decentralized compute has been handed, because it proves the market pays for efficient inference, and efficient inference is what these networks are built to supply.
The honest caveats
Two things could weaken the thesis, and they deserve stating. First, decentralized compute still faces genuine hurdles on latency, reliability, security, and the verifiability of remote computation — an enterprise running production inference needs guarantees that an anonymous GPU network has not fully solved. Verifiable inference, where a network can cryptographically prove it ran the model you asked for, is the missing primitive, and it is not finished. Second, hyperscalers will cut inference prices aggressively to defend the workload, and their integration and reliability advantages are real. The decentralized cost advantage has to survive that response.
But the direction of the evidence is one-way. The company winning on revenue is the one spending less, the workload that is growing is the one that distributes well, and the price of the centralized alternative is rising under a hardware shortage. Those three facts point at the same conclusion, and they were not arranged to. That is what makes the signal credible rather than convenient.
Frequently asked questions
Did Anthropic really overtake OpenAI in revenue? Yes, in annualized run-rate terms as of April 2026. Anthropic reached roughly $30 billion annualized against OpenAI’s approximately $24 billion, having grown from about $1 billion just fifteen months earlier. Epoch AI had forecast the crossover for around August 2026, so it arrived ahead of schedule. The figures come from company disclosures and reporting by outlets including The Information and Bloomberg, aggregated by Epoch AI and others. Revenue run rate is a snapshot, not audited annual revenue, but the gap and the growth trajectory are consistent across independent sources, which is why the crossover is treated as real rather than a one-off.
How can Anthropic make more money while spending far less on training? Its revenue is roughly 85% enterprise and developer customers who pay for high-value inference wired into real workflows, rather than a mass consumer product where most users pay nothing. Enterprises buy reliable, affordable model access and turn it into business value, which supports strong pricing. Because the revenue does not depend on subsidizing a free consumer base, Anthropic does not need to win every frontier training race to monetize — it can spend an estimated 4x less on training (peaking near $30 billion in 2028 versus OpenAI’s $121 billion) and still out-earn on the strength of profitable inference demand.
Why is this good news for decentralized compute? Because it shifts the industry’s binding constraint from building giant training clusters to supplying cheap, scalable inference — and inference is the workload that distributes well across a network of GPUs. If capital-efficient inference is what converts to revenue, then networks that undercut hyperscaler inference pricing gain a genuine buyer rather than an ideological one. Projects like Render, Akash, and io.net are built to supply exactly that market-priced capacity. The shift does not guarantee they win, but it hands decentralized compute the demand-side argument it has lacked, grounded in where the revenue is actually going.
Which decentralized compute tokens are most relevant to this thesis? Render (RNDR) and Akash (AKT) run live GPU marketplaces that already undercut centralized cloud pricing on comparable hardware, with Render extending toward AI inference. io.net (IO) aggregates GPUs into ML-focused clusters aimed at the same cost gap. Bittensor (TAO) is a higher-risk bet that decentralizes the model-serving layer itself through token incentives. None is a guaranteed winner, and all carry the reliability, latency, and verifiability risks discussed above. The thesis is about the category gaining a demand case, not a recommendation to buy any specific token.
What is the biggest risk to this argument? That hyperscalers defend inference aggressively on price and integration while decentralized networks fail to solve verifiability — proving cryptographically that a remote GPU actually ran the model requested. Enterprises need reliability and security guarantees that anonymous GPU networks have not fully delivered, and centralized providers will cut inference prices to keep the workload. If verifiable inference does not mature and the cost advantage erodes under hyperscaler price competition, decentralized compute could stay a niche. The thesis rests on cost and workload structure favoring distribution; both the cost gap and the verifiability problem are the variables to watch.
Sources
- SaaStr — Anthropic Just Passed OpenAI in Revenue While Spending 4x Less to Train Their Models
- Trending Topics — Anthropic Overtakes OpenAI in Revenue, Hitting $30 Billion Run Rate
- Epoch AI — Anthropic could surpass OpenAI in annualized revenue by mid-2026
- Forbes — OpenAI and Anthropic Are Testing Two Very Different AI Business Models
- Render Network — decentralized GPU marketplace
- Akash Network — decentralized compute marketplace
- io.net — decentralized GPU clusters for machine learning
- Bittensor — incentive network for machine intelligence
What Anthropic’s Revenue Comparison With OpenAI Reveals About the Limits of Headline Numbers in AI Market Analysis
“Passed” is doing a lot of work in this headline. The probabilistic question is: what is the uncertainty range around both revenue figures, and how confident can we be that the comparison is comparing equivalent things? OpenAI’s revenue has been variously reported by the company, by investors in fundraising contexts, and by journalists citing unnamed sources — each with different methodological definitions of what counts as revenue. Anthropic’s revenue figure is similarly reported from non-public sources. When two numbers with significant uncertainty ranges are compared, the probability that the comparison is actually correct is lower than the headline’s precision implies. The honest framing is a range estimate with explicit uncertainty, not a point comparison presented as a settled fact.
The “4x less training spend” framing raises a second measurement problem. Training spend and revenue exist in different time frames. Anthropic’s current revenue reflects products built on models trained in prior periods; the training spend that generated those capabilities was incurred earlier. Comparing current revenue to current or cumulative training spend conflates a flow metric with a cost metric that spans multiple periods. The implicit efficiency claim — that Anthropic has found a more capital-efficient path to revenue — may be directionally correct, but the metric pair chosen to illustrate it does not establish the claim cleanly. A fair comparison would require knowing the training spend attributable to each company’s current production models, divided by the revenue those specific models generate.
The market analysis implication of the revenue comparison, if taken at face value, is that the AI model competitive landscape is more balanced than the ChatGPT brand dominance story implies. A world where Anthropic has genuinely higher revenue than OpenAI is a world where Claude’s enterprise adoption has developed faster and more durably than consumer ChatGPT usage would suggest. That would be a significant finding: enterprise buyers are choosing Claude over GPT-4o at a rate the consumer market does not reflect. But that conclusion requires accepting the headline numbers’ precision. The probabilistic assessment is to hold that conclusion with meaningful uncertainty, weight it lightly until either company discloses audited revenue figures, and watch the next fundraising round’s valuation — which is the data point most likely to reveal which revenue figure institutional investors actually believe.
