FundingRounds, leads & terms
Baseten CEO calls inference AI's 'final market' as reported valuation talks reach $26B–$30B
Inference provider Baseten went from a $5B valuation in January 2026 to reported talks at $26B–$30B in October. CEO Tuhin Srivastava argues that inference will become AI's largest and most durable market as companies run their own post-trained models.
From the The Generalist article of November 17, 2026 and cited sources: Baseten's valuation has risen quickly. [3] [6] [8] [4]
SOURCES[3] Baseten: Announcing our Series F baseten.co[6] Verdict: Baseten secures $1.5bn Series F verdict.co.uk[8] Bloomberg: Baseten set for $30B valuation bloomberg.com
AI-assisted, not human-edited. Methodology Corrections Who owns this site
- Original source
- The Generalist article, independent perspective
- Brief by
- AT Pulse research desk AI-assisted; publisher and contact on the desk page
- When
- November 17, 2026 article published
- Covered
- Nov 17 edition
- Evidence
- 9 sources across 8 sites
- 1Why inference is harder than you thinkgeneralist.com
- 2Linked video: Why inference is harder than you thinkyoutube.com
- 3Baseten: Announcing our Series Fbaseten.co
- 4TokenPost: Baseten seeks $2Btokenpost.com
- 5Podwise episode datepodwise.ai
- 6Verdict: Baseten secures $1.5bn Series Fverdict.co.uk
RisingHow it’s measured
YouTube video · 13.1K views · 488 likes
Reach blends public engagement (views, likes, comments, shares, stars) from the platforms where the story or its original source lives: YouTube, Hacker News, GitHub, Substack and similar. A video and the article written from its transcript share one score, taken from the video. Platforms with no data are left out rather than counted as zero, and the age of a story never lowers its score.
Baseten's valuation has risen quickly. It raised $300M at $5B in January 2026, then $1.5B at $13B in June. That Series F closed in two tranches priced at $13B and $11B. In October, it was reported to be raising about $2B at $26B–$30B. [3] [6] [8] [4]
Srivastava calls inference AI's final market and puts inference spending at roughly $2T by 2030. He also says the company has underforecast its own demand every year since 2023, and that its January 2025 forecast came in about 4x below year-end actuals. [1]
Srivastava describes a typical path for customers. They start with closed APIs, move to dedicated endpoints for latency, isolation and cost, and larger firms then self-host in their own cloud. In his 'owned intelligence' loop, companies deploy open-weight models, collect feedback, post-train and redeploy. [2] [1] [6]
Inference needs different things from training: compute located close to users, high reliability, and repeatable workflows such as versioning and cold starts. Speed gains come from newer GPUs, continuous batching, speculative decoding, smaller distilled models and kernel-level tuning on Nvidia's TensorRT-LLM. [2]
The funding rounds and investors are documented, but the growth multiples and the $2T forecast come from Baseten itself.
Read the full assessment
If inference spending grows as claimed, choosing between shared, dedicated and self-hosted serving becomes a major cost and strategy decision.
Executive brief
Baseten's own forecasts keep falling short of reality. CEO Tuhin Srivastava says the January 2025 forecast missed year-end actuals by about 4x, and the company has misforecast every year since 2023 (The Generalist). Its valuation is moving just as fast. Baseten announced $1.5B at $13B in June 2026 (Baseten). By October, unnamed sources said it was raising about $2B at $26B–$30B (TokenPost, citing Bloomberg). Srivastava's main argument in The Generalist interview is that inference is AI's "final market," and he gives a rough estimate of about $2T in inference spending by 2030. The figure is his own and nothing in the reviewed sources supports it independently.
What changed and event timeline
No Priors episode on inference speed
This is the video linked from the article, and it dates from 2024. In it, Srivastava says customers are moving from closed APIs to dedicated endpoints and then to self-hosting, and that buying inference beats building it (;).
Second No Priors interview
A third-party summary reports that 95% of the tokens Baseten serves come from post-trained models and that the company grew revenue 30x. Both figures are self-reported ().
$1.5B Series F at $13B
Altimeter, Conviction and Spark led, with Sands Capital and Wellington as co-leads (). Verdict reports the round closed in two tranches, at $13B and $11B ().
Reported $26B–$30B round
Bloomberg, citing unnamed sources, reported about $2B in two tranches, with Sands Capital expected to lead (;).
The Generalist interview published
He also argues that hyperscalers split inference across four internal teams, and that "owned intelligence" works as a moat ().
Capabilities and access
Read the full section
- Products: dedicated inference, model APIs, training, sandboxes, and Grid, a model access-control product (Baseten).
- Named customers: Cursor, Notion, Harvey, HubSpot, OpenEvidence, Abridge and Decagon (Baseten).
- Footprint: the sources disagree. The Generalist says 20 clouds and 90 regions. Verdict, quoting Baseten, says 18 cloud providers and 87 clusters, serving more than 1B inference calls a day (Verdict).
- Deployment options: shared endpoints, dedicated compute, or a self-hosted license that runs in the customer's own cloud (video).
Technical analysis for researchers and developers
The 2024 episode describes how Baseten approaches the problem: None of the reviewed sources includes a published benchmark methodology.
Read the full section
The 2024 episode describes how Baseten approaches the problem:
- Inference vs. training: inference needs compute located close to users, strong reliability, and repeatable workflows such as versioning, CI/CD and cold starts. Training depends mostly on interconnect between GPUs (video).
- Where speed comes from: better hardware (H100 over A100), continuous and dynamic batching, speculative decoding, and smaller distilled models (video). Baseten built on and contributed to Nvidia's TensorRT-LLM, and Srivastava notes that much of the optimization happens at the kernel level (video).
- Non-Nvidia chips: he is skeptical that CUDA-style workloads port easily to AMD, because customers change drivers, PyTorch versions and Docker images (video).
None of the reviewed sources includes a published benchmark methodology.
Claims and evidence
Read the full section
- Round size, valuation and investors: announced by Baseten and repeated by Verdict (Baseten; Verdict). No filing confirms them.
- $26B–$30B round: comes only from unnamed sources (TokenPost).
- Growth figures: 20x revenue and 40x inference volume growth are self-reported (Baseten). So are the 30x growth, 400% net dollar retention and more-than-$1B revenue trajectory (digest).
- $2T by 2030 and the 4x forecast miss: the CEO's own estimates, with no corroboration (The Generalist).
- Disclosure: Sarah Guo co-hosts No Priors and founded Conviction, which co-led the Series F. In 2024 she praised Baseten's Artificial Analysis benchmark results on air (video).
Context and prior work
- Competitors: Modal and Baseten were in funding talks at the same time (Bloomberg). Fireworks AI reached a $17.5B valuation in July 2026 (TokenPost).
- Specialized hardware: The Generalist's resources list Etched, Cerebras and Nvidia's Groq 3 LPX (The Generalist).
- Same thesis in 2024: throughput was rising from about 90 to 300–400 tokens per second, and Srivastava expected that performance to become a commodity eventually (video).
Read the full section
- Competitors: Modal and Baseten were in funding talks at the same time (Bloomberg). Fireworks AI reached a $17.5B valuation in July 2026 (TokenPost).
- Specialized hardware: The Generalist's resources list Etched, Cerebras and Nvidia's Groq 3 LPX (The Generalist).
- Same thesis in 2024: throughput was rising from about 90 to 300–400 tokens per second, and Srivastava expected that performance to become a commodity eventually (video).
Limitations, safety and contested findings
- Footprint numbers conflict: 20 clouds and 90 regions vs. 18 providers and 87 clusters.
- Valuation is not a single number: the two-tranche structure ($13B and $11B) makes the headline valuation softer than it looks (Verdict).
- Compute supply: in 2024, customers spent three to six weeks negotiating for premium GPUs (video).
Read the full section
- Footprint numbers conflict: 20 clouds and 90 regions vs. 18 providers and 87 clusters.
- Valuation is not a single number: the two-tranche structure ($13B and $11B) makes the headline valuation softer than it looks (Verdict).
- Compute supply: in 2024, customers spent three to six weeks negotiating for premium GPUs (video). By 2026, new GB200 capacity reportedly required 3–5-year contracts with 20–30% prepaid, which makes cost of capital the binding constraint (digest).
- Where inference may run: Srivastava himself says inference moving onto local devices would hurt cloud providers (video).
Business and practitioner implications
- Inference can be a major cost line. One customer's CTO said inference was their second-largest expense after payroll (video).
- Buying can beat building. This is an anecdote from the vendor.
- The "owned intelligence" loop: deploy open-weight models, collect feedback, post-train, then redeploy (The Generalist).
Read the full section
- Inference can be a major cost line. One customer's CTO said inference was their second-largest expense after payroll (video).
- Buying can beat building. A team with four infrastructure engineers that had spent two years on its own stack moved all its workloads to Baseten in 36 hours (video). This is an anecdote from the vendor.
- The "owned intelligence" loop: deploy open-weight models, collect feedback, post-train, then redeploy (The Generalist). Baseten says leading application companies spend 30–50% of their model budget on custom models, a figure it reports itself (Verdict).
The source trail.
Sources (9)
Why inference is harder than you think
Feed excerpt only. A transcript of the video linked from this page is supplied (linked video auto captions; https://www.youtube.com/watch?v=oe3OSlsf4YA); automatic text may contain errors.
generalist.comLinked video: Why inference is harder than you think
linked video auto captions
www.youtube.com