Baseten grew revenue 10x by selling independence, not GPUs

1. Three layers of problems — ▶ 2:08

   Inference is serving models, and Baseten splits it into three layers. At the bottom is heterogeneous compute, pulled from 20 clouds and 90 regions and hidden from customers, because nobody can buy 1,000 top-end GPUs on the open market right now. In the middle is the core inference stack: speed, reliability and the right numerics. On top are inference-adjacent primitives such as post-training and RL, sandboxes for executing model-generated code, evals and routing. Closed frontier providers bundle all of this. With open or post-trained models, the customer must assemble it behind an API.

2. Specialist versus hyperscaler — ▶ 5:26

   His market arithmetic: AI spend in five years of $2-10 trillion at roughly a 50% gross margin implies $1-5 trillion of inference spend. That means $10-50 trillion of market cap split among hyperscalers, closed labs and inference providers. Hyperscalers will have inference products but are also strong partners. The case for Baseten is focus: "we do one thing, they do many." When an inference provider is down, the customer's product is down, so buyers pick whoever moves fastest and pays the most attention. Off-the-shelf inference isn't dialed in for production scale.

3. Compute is not the moat — ▶ 9:07

   The criticism that inference clouds just resell GPUs at thin margins is partly right, and he says so. Compute is necessary, but until the software loop works it is reselling. The loop is owned intelligence: run open and closed models, evaluate each release, post-train in your own environments, execute code in sandboxes, ship a better model, repeat. Companies want it for cost control, data sovereignty and "credible permanence", meaning models can't vanish on you. Owning the loop protects margins, so customers will pay Baseten a lot of that margin.

4. Adoption turned in December — ▶ 14:15

   Model capability jumped in December, with another shift in the last month. Over 6-7 months coding models got good, a compute shortage hit and open-weight models improved release after release. Earnings calls now judge companies on how much they spend on AI. That pushes them to rethink long-term cost strategy: how to put AI in more places without blowing up the company. He says this has taken over the narrative and pushed adoption further, and that Baseten customers are growing so fast he gets paged mid-interview.

5. Serving was the least important facet — ▶ 15:45

   His path ran from Australia to infrastructure finance in New York (toll roads, bridges, airports), then a Boston lab using machine learning to diagnose neuromuscular diseases from 2012. In 2019 he started Baseten with friends Phil and Amir, betting on picks and shovels for machine learning with no commercial proof. Model serving looked like the least important facet of their infrastructure company. After the market accelerated in 2022, it became the entire company. He says application-layer work needs domain expertise they lacked.

6. Open as majority of tokens — ▶ 19:11

   The valuation went from $5 billion in January 2026 to $13 billion in June. Inside Baseten the question moved from whether open-weight models would matter to whether they might be a majority of token volume. His future is mixed intelligence: closed frontier, open weights and post-trained models, with custom models taking a share. The economics are simple. If a 2-month-old model handles 90% of your tasks at a fraction of the cost and you keep your data, that is "a no-brainer" and very rational.

7. DeepSeek fear was capability fear — ▶ 20:31

   After DeepSeek V3 launched at Christmas 2024, he called about 10 enterprises. At least 90% said they would never use it in production because they didn't know its origin or how it worked. Eighteen months on, with GLM, Kimi, DeepSeek, Nvidia's Nemotron family and American models like Inkling available, every enterprise he meets is keen to run open models. His reading is that the fear was driven by capability, not by real analysis. He expects strong American open-weight models in weeks to quarters, which would wash away the "are we falling behind" narrative.

8. Intellectually honest dependency — ▶ 24:19

   Asked whether Baseten must ride open-weight growth, he says "the intellectually honest answer is yes." He softens it: open models needn't exist, but customers must want to own their intelligence, and open weights are an accelerant. Baseten launched Base Labs, a research arm built from an acquired team. Today it post-trains with customer data and sets up learning loops, and it does not do large-scale pre-training. It is part of Nvidia's Nemotron coalition. If it gains a differentiated edge and the compute allows, it will train its own models.

9. Vertical for destiny, not just value — ▶ 30:07

   He gives two reasons to own more of the stack. Going up adds customer value and captures more of the learning loop. Going down is about control: he expects Baseten to be one of the largest individual users of compute on the planet within 12-24 months, "the labs and then Base 10." Current scale: 40-50 trillion tokens a day, token volume up 40x year on year, thousands of customers, and by his recollection revenue up roughly 10x in 12 months. He says tokens from your apps today have likely passed through Baseten.

10. Sandboxes belong with inference — ▶ 33:53

    The Blacksole acquisition rests on the view that inference and sandboxes are extensions of one problem: where models run is where the execution environment should be. Sandboxes mostly run on CPUs, which raises the question of co-locating CPUs and GPUs and sends him back down to heterogeneous compute. Agents like Instinct and Town are first examples. He sees current penetration at 1-2% and inference becoming 100,000x bigger, and says that will need more data centers, more power and aggressive inference optimization.

11. Last market, agent customers — ▶ 43:21

    The "last market post-AGI" claim means only this: without AGI it is the biggest market ever, and with AGI models running in loops are the only workload left. He disclaims any AGI expertise. Agents are already Baseten customers. They deploy models via its MCP server and skill, pick which model to use and debug live incidents with access to code and architecture. He won't delegate large-impact production actions to them. The human skill shifts to deciding how to interact with the data.

Compute is the commodity. The product is the loop. Srivastava is blunt about his own business here: until the software layer works, a provider with only GPUs is reselling capacity. The loop is what changes the economics.

The loop has a specific shape. A company runs open-weight, post-trained and closed frontier models side by side. It evaluates each new release. It post-trains on its own environments and its own user feedback. It gives models sandboxes to execute code. Then it serves a better model and starts again. Srivastava calls the result owned intelligence, and he gives three reasons a company wants it: cost control, data sovereignty and what he calls credible permanence, meaning models cannot disappear on you one day. The data and the model improvements become core IP. Feed them into someone else's loop and you are improving the lab that supplies you. Run your own and you protect your margins. Because Baseten supplies the loop, customers will pay it a large share of those protected margins. It is, in his phrase, "making them independent again." The interviewer notes this is a different meaning of continual learning from the one frontier labs use, where models improve themselves. Srivastava's version is about who controls the improvement.

That also answers the obvious question of why not AWS, Azure or Google Cloud. He does not treat hyperscalers as enemies. They will have inference products, and they are partners with strong relationships to Baseten. His case is focus. They do many things; Baseten does a few. When an inference provider is down, the customer's product is down, so buyers go to whoever will move fastest and give them the most attention. The prize explains the crowding. His back-of-envelope: AI spend in five years of somewhere between $2 trillion and $10 trillion, at about a 50% gross margin, means $1-5 trillion of inference spend. Divided among hyperscalers, closed labs and inference providers, that is $10-50 trillion of market cap. He admits the headline number does not matter. The point is that any version of it is enormous.

The loop only pays if enterprises are willing to run models they can own, which makes open weights the load-bearing assumption. His picture of the future is mixed intelligence: closed frontier, open weights and post-trained models together, with custom models taking a share. The logic is cost. AI could already solve a lot two months ago. If a 2-month-old model handles 90% of your tasks at a fraction of the price, and your data stays yours, that is "a no-brainer." Inside Baseten the question shifted in six months from whether open weights would matter to whether they could be a majority of token volume. That shift sits behind a valuation that went from $5 billion in January 2026 to $13 billion in June. The operating numbers back it: 40-50 trillion tokens a day, token volume up 40x year on year, thousands of customers and, by his recollection, revenue up roughly 10x in 12 months.

Enterprise attitudes moved faster than most people expected. When DeepSeek V3 landed at Christmas 2024, he called about 10 enterprises he had met over the years. At least 90% said they would never use it in production. They did not know where it came from or how it worked. Eighteen months later, with GLM, Kimi, DeepSeek, Nvidia's Nemotron family and American models like Inkling on the table, he says every enterprise he talks to wants to run open models and few argue about origin. His reading is that the fear was really a capability gap. It disappeared the moment the models were good enough. The same reasoning underlies his prediction that strong American open-weight models arrive within weeks to quarters, and that the current anxiety about falling behind China will wash away with them.

The obvious objection is that Baseten has tied its growth to something it does not control. Asked directly, he concedes it: "the intellectually honest answer is yes." He softens the point. Open weights need not exist in the abstract; customers need to want ownership, and open models accelerate that. But the dependency is real, and it explains the rest of his strategy. Supply is tight. Nobody can buy 1,000 top-end GPUs on the open market, so Baseten stitches together 20 clouds and 90 regions. He moves up the stack to deepen customer value and capture more of the loop. He moves down because he expects Baseten to be one of the largest users of compute on the planet within 12-24 months, and he wants some control over what sits upstream and downstream. Base Labs, the research arm, post-trains today. Large-scale pre-training is conditional on a differentiated edge and enough compute.

Agents are the next source of load, and they pull Baseten further down the stack. The Blacksole acquisition rests on the view that inference and execution environments are one problem: where models run is where code should execute. Sandboxes mostly run on CPUs, so the question becomes whether CPUs and GPUs should be co-located, which feeds back into heterogeneous compute. He points to consumer agent products like Instinct and Town as first examples and argues that the productivity gains already visible in healthcare, legal and customer support must reach consumers. Penetration, he says, feels like 1-2%. Agents are also already customers. They deploy models on Baseten, choose which model to use and debug incidents with access to the code and architecture. Humans keep control of large-impact production actions, and the human skill shifts to deciding what data agents interact with.

Several things stay unresolved. His answer on scaling is that more data centers and more power are needed and that inference costs will keep falling, but he offers no figures. He admits data-center backlash slows Baseten, and he frames the fix as better education from an industry that has failed to bring people along. When pressed on existential risk, he talks about guardrails and listening rather than anything that constrains his own roadmap. The "last market post-AGI" line turns out to be modest: he disclaims expertise on AGI and means only that there will be a great deal of AI to serve. The deepest tension is the one his pitch creates. Ownership is the product, so a company that escapes the labs by running its learning loop on Baseten has chosen a new dependency. Whether customers see that as independence depends on how much of the protected margin Baseten ends up keeping.