AI Inference Valuations Double in 90 Days: Baseten $26B, Modal $15B

AI inference valuations are repricing faster than revenue. Baseten is in talks at roughly $26 billion, double its $13 billion June mark, Axios reported on September 22. Bloomberg reported on September 23 that Modal is negotiating at about $15 billion, up from the $4.65 billion it announced in May. Neither round has closed. Both imply multiples above 40x annualized revenue.

What are Modal and Baseten actually raising at?

Two numbers, both from reporting rather than filings. Axios reported on September 22 that Baseten is in discussions at about $26 billion. Bloomberg reported a day later that Modal is in talks near $15 billion. The amount of capital either company hopes to raise has not been disclosed.

That distinction matters. These are proposed valuations from live negotiations, not completed financings. Terms move.

What is documented is the trail behind them. Modal published its own Series C announcement on May 21, 2026: $355 million at a $4.65 billion post-money valuation, led by General Catalyst and Redpoint, with Menlo, Bain Capital Ventures and Accel joining. Baseten confirmed $300 million at a $5 billion valuation in January 2026 via Business Wire, led by IVP and CapitalG with Nvidia participating.

The valuation ladder, round by round

Company Date Round Amount Valuation
Modal Sep 2025 Series B $87M $1.1B
Modal May 21, 2026 Series C $355M $4.65B
Modal Sep 2026 In talks Undisclosed ~$15B
Baseten Sep 2025 Series D $150M $2.15B
Baseten Jan 2026 Series E $300M $5B
Baseten Jun 2026 Series F $1.5B $13B
Baseten Sep 2026 In talks Undisclosed ~$26B

Baseten has gone from $2.15 billion to a reported $26 billion in twelve months. Modal has moved 13.6x in the same window. Neither company sells a model. Both sell the machinery that runs one.

Why are AI inference valuations doubling in 90 days?

Because inference revenue is compounding at rates venture math has no template for, and because 2026 is the first year enterprises spend more running models than training them. Investors are underwriting a usage curve, not a product cycle. The bet is that token volume keeps outrunning price cuts.

The revenue actually behind the numbers

Baseten reached roughly $600 million in annualized revenue by March 2026, according to Sacra, up from about $200 million in December 2025 and roughly $30 million a year earlier. That is a 20x year-over-year move.

Modal’s own Series C post put the company above $300 million in annualized revenue in May, described as fivefold growth since its September Series B. Sacra puts Modal’s end-2025 figure near $119 million.

Modal also disclosed something more interesting than a revenue line: over one billion sandboxes launched on the platform, with sandboxes driving more than a third of revenue. Agent workloads, not chatbots, are the volume story.

  • Baseten customers named publicly: Cursor, Notion, Abridge, Clay.
  • Modal customers named in its Series C post: DoorDash, Cognition, Decagon, Suno, Ramp, Physical Intelligence, Reducto.
  • Overlap: both serve the coding-agent and voice-agent cohort — the fastest-burning compute buyers in the market.

What Gartner says about the inference pivot

Gartner forecast on August 10 that worldwide AI-optimized IaaS spending will reach $42.276 billion in 2026, up 96.4% from $21.529 billion in 2025, and $66.143 billion in 2027.

Inside that, Gartner splits 2026 at $23.3 billion for inference against $19 billion for training — 55% versus 45%. It is the first year inference takes the larger share.

Hardeep Singh, Gartner’s senior principal research analyst, framed the shift plainly: as organizations move “from model development to production-scale deployment,” domain-specific models get wired into customer-facing systems “requiring continuous, real-time execution.”

That is the thesis in one sentence. Training is a project. Inference is a utility bill.

How do these multiples compare to Together AI?

Badly, for Modal and Baseten. Together AI closed $800 million at an $8.3 billion post-money valuation on July 1, led by Aramco Ventures, with Nvidia, Vista Equity Partners and General Catalyst participating. It disclosed $1.15 billion in annual bookings — the largest revenue base of the three.

Run the arithmetic against the reported marks:

Company Valuation Revenue basis Implied multiple
Modal ~$15B (in talks) $300M annualized, May 2026 ~50x
Baseten ~$26B (in talks) $600M annualized, Mar 2026 ~43x
Together AI $8.3B (closed) $1.15B bookings ~7.2x

Together AI has nearly twice Baseten’s disclosed revenue at roughly a third of the valuation. Some of that gap is stale denominators — Baseten’s March figure is six months old and both companies are growing fast. Some of it is not.

The skeptical read: a closed round at 7x and a rumored round at 43x are not pricing the same asset. They are pricing different assumptions about who keeps the gross margin when Nvidia capacity stops being scarce. Inference is, structurally, a business that rents someone else’s silicon and resells it with software on top. That software layer is real. It is also the thing hyperscalers are building in-house.

Who profits from the inference land grab?

Nvidia, first and most reliably. It sits on the cap table of both Baseten and Together AI while selling the GPUs all three resell. Every markup in this sector is downstream of hardware Nvidia already sold.

After that, the winners thin out:

  1. Early-stage funds holding seed and Series A paper. Baseten’s seed was $2.5 million led by First Round; its Series A was $13.5 million led by Sequoia. At $26 billion those are generational marks — on paper.
  2. Crossover funds writing the late rounds. Altimeter, Conviction and Spark led Baseten’s $1.5 billion Series F at $13 billion in June. A $26 billion mark doubles that position in one quarter, unmarked by any exit.
  3. Open-weight model labs. Both platforms specialize in running open-source models. Cheap open weights are the demand engine; Together AI claims customer cost savings of 6x to 60x against closed-model pricing.
  4. Public-market investors — indirectly and later. None of these three is listed. Exposure runs through Nvidia, the hyperscalers and the power complex.

Why this matters for the wider AI market

Inference is where AI stops being capex and starts being revenue. The valuations attached to Modal and Baseten are the private market’s estimate of how much of that revenue the middle layer gets to keep.

Three implications worth holding onto.

First, the marks are moving on secondary reporting, not closed paper. Both stories originated as scoops. Rounds reported at $15 billion and $26 billion have repriced downward before.

Second, this is the same velocity visible elsewhere in the stack — see our coverage of Snorkel and Micro1 marking up on the same day and Crusoe’s 3.1x markup in eleven months. Compression of the venture clock is now a sector-wide feature, not a company-specific one.

Third, none of it has cleared a public market. AI IPO volume in 2026 was $90 billion and 83% of it was a single company. Private marks at 43x are untested by any liquid bid. The financing structures underneath — including roughly $300 billion in off-balance-sheet AI debt — add leverage to the same wager.

This post is reporting and analysis, not financial advice.

Frequently asked questions about AI inference valuations

Have the Modal and Baseten rounds closed?

No. Both were reported as active discussions — Axios on September 22 for Baseten, Bloomberg on September 23 for Modal. Neither company has confirmed final terms, and the raise amounts are undisclosed.

What is AI inference, in business terms?

Running a trained model to produce outputs. Training is a one-time capital cost; inference is a recurring operating cost that scales with usage. Gartner expects $23.3 billion of inference spending on AI-optimized IaaS in 2026.

How much revenue does Baseten have?

Roughly $600 million annualized as of March 2026, per Sacra, up from about $200 million in December 2025. Baseten has not published an updated figure alongside the $26 billion report.

What was Modal’s last confirmed valuation?

$4.65 billion post-money, from the $355 million Series C the company announced on May 21, 2026, led by General Catalyst and Redpoint.

Is Nvidia invested in these companies?

Nvidia participated in Baseten’s $300 million Series E at a $5 billion valuation and in Together AI’s $800 million round at $8.3 billion. It is both supplier and shareholder.

Why is Together AI valued lower with more revenue?

Its $8.3 billion is a closed July round; the Modal and Baseten figures are rumored September marks. Private valuations also reflect growth rate and round timing, not revenue alone.

Can retail investors buy into AI inference platforms?

Not directly — all three are private. Exposure is indirect, via chip suppliers, cloud providers and power infrastructure. This is not investment advice.

The bottom line

AI inference valuations have decoupled from the revenue multiples the rest of software trades on, and the gap is widest exactly where the numbers are least confirmed. Baseten at $26 billion and Modal at $15 billion are reported marks on undisclosed raises, priced at 43x and 50x six-month-old revenue figures.

The demand is real: Gartner’s 96.4% growth in AI-optimized IaaS and the first-ever inference majority are not sentiment, they are budget. What is unproven is margin durability in a layer that resells Nvidia capacity to customers who can switch in an afternoon.

Watch three things next. Whether either round closes at the reported number. Whether Baseten publishes revenue above $600 million when it does. And whether Together AI’s $8.3 billion mark gets dragged up — or whether the 6x spread between closed and rumored pricing turns out to be the market telling the truth. For more on how private AI marks are set, see our analysis of Cohere’s $20 billion valuation at 83x ARR.

Sources


Comments

Leave a Reply

Discover more from Wealth Engine

Subscribe now to keep reading and get access to the full archive.

Continue reading