Booming AI Revenues Boost Inference Startups to Decacorn Status

---

1. Key Themes

Theme 1: Inference is the New Epicenter of AI Investment

Enterprise AI adoption is driving explosive demand for inference compute, transforming a once-questioned sector into one of the hottest investment categories in tech. Multiple companies are raising at decacorn valuations in rapid succession.

"Inference computing, or the process of running a trained machine learning model on new data to generate predictions or outputs, is suddenly blowing up alongside the heightened demand for AI tools from enterprises."

Revenue growth substantiates the thesis: Fireworks surged from $250M to $800M ARR in roughly seven months; Baseten's ARR reportedly jumped from $200M to $600M in a single quarter; Modal crossed $300M in ARR.

---

Theme 2: Revenue Momentum Is Overriding Margin Scrutiny — For Now

VCs are aggressively deploying capital into inference startups based on top-line growth, setting aside traditional concerns about unit economics. The market is rewarding scale over profitability.

"The revenue momentum for all of these companies is hard to deny," said Deedy Das, partner at Menlo Ventures, citing that many were growing at multiples "on a $100 million-plus baseline" in the first half of 2026.

"It seems like VCs are just doing a revenue multiple and are assuming the margin doesn't matter," said one skeptical investor.

---

Theme 3: The Structural Vulnerability of Leased-Compute Inference Players

Inference startups that lease rather than own compute face a compounding disadvantage — rising GPU costs, competition from hyperscalers, and supply pressure from the labs themselves.

"Baseten, Fireworks AI, and Modal all just lease capacity, unlike neoclouds such as Lambda and Crusoe which provide inference while also owning the chip stack. Additionally, they're also competing with the labs themselves for compute allocations, as OpenAI and Anthropic are gobbling up more chip capacity for their own training and development."

---

Theme 4: Coding and Enterprise Workloads as the Wedge for Inference Demand

Coding assistants have emerged as a primary enterprise use case accelerating inference demand, with Cursor explicitly named as a major customer anchor.

"Coding assistants have been one area where enterprise adoption generated much more need for better inference, and Fireworks AI in particular has depended on Cursor as a major customer. But even just running LLM queries on internal company data requires inference capacity, so the market is theoretically set to grow much larger as more businesses adopt AI tools."

---

2. Contrarian Perspectives

The Inference Stack May Commoditize Before These Valuations Are Justified

The article surfaces a credible bear case: inference providers have largely similar offerings, compete on price, and are at risk of being squeezed from above (labs) and below (hyperscalers). Differentiation is thin and narrowing.

"They also have fairly similar product offerings, bringing the risk of customers switching back and forth depending on who offers the best price of the moment... as both expand into other parts of the stack besides inference, like fine-tuning, the risk of commoditization is high."

The companies' lack of owned compute makes margin expansion structurally difficult at exactly the moment when they are being valued as durable businesses.

---

Owning the Compute Layer May Be the Only Defensible Position

Against the consensus of backing inference-as-a-service, the article implicitly points to neoclouds (Lambda, Crusoe) as structurally better-positioned because they own the chip stack rather than lease it — a distinction that will matter more as pricing pressure intensifies.

"Unlike neoclouds such as Lambda and Crusoe which provide inference while also owning the chip stack" [inference startups] "just lease capacity."

---

Rapid Re-Raises Signal Froth, Not Just Momentum

Baseten raising up to $1 billion just four months after its previous round is either a sign of extraordinary business velocity — or of a market willing to paper over structural risks with fresh capital. The speed of re-raising is notable even by AI standards.

"Inference provider Baseten is raising up to $1 billion in new funding just 4 months after closing its previous round, and is looking for an $11 billion valuation."

---

3. Companies Identified

---

4. People Identified

---

5. Operating Insights

Concentration Risk in Enterprise Customers Is Real — Even at $800M ARR

Fireworks AI's heavy dependence on a single customer (Cursor) at massive scale is a cautionary tale for operators: revenue growth can mask dangerous customer concentration. Diversification across enterprise verticals is essential before large fundraising rounds amplify the risk.

"Fireworks AI in particular has depended on Cursor as a major customer."

---

Re-Raising Before Capital Is Needed Is Now a Strategic Offensive Move

Baseten's decision to raise up to $1B just four months after closing a prior round reflects a deliberate strategy to lock in capital and valuation momentum while market conditions are favorable — rather than waiting until cash is needed. In an environment where revenue multiples are high and momentum is fundable, operators should treat fundraising as a proactive competitive lever.

"Inference provider Baseten is raising up to $1 billion in new funding just 4 months after closing its previous round."

---

Expanding Up the Stack Is Necessary but Accelerates Commoditization Risk

Both Baseten and Fireworks are moving beyond inference into fine-tuning and evaluation tools to build more defensibility. But this expansion is converging their product roadmaps, which could undermine pricing power across both companies. Operators expanding product scope should be deliberate about avoiding feature parity with direct competitors.

"As both expand into other parts of the stack besides inference, like fine-tuning, the risk of commoditization is high."

---

6. Overlooked Insights

On-Device AI Could Be a Quiet Threat to Inference Cloud Revenue

Apple's push for AI models that run locally on devices — if successful at scale — represents a potential demand headwind for cloud inference providers. If inference moves to the edge for a significant portion of use cases, the TAM for companies like Baseten and Fireworks could face structural compression.

"Apple makes a big push for AI models that can work locally on its devices ahead of a new Siri."

---

Large Enterprises Are Building Proprietary AI Infrastructure In-House

Kirkland & Ellis allocating hundreds of millions to build its own AI legal tools signals a broader trend: large, high-margin professional services firms may choose to own their AI tooling rather than rely on third-party inference APIs. If this pattern spreads, it could reduce the addressable enterprise market for inference-as-a-service over time.

"Big law firm Kirkland and Ellis allocates hundreds of millions to make its own AI legal tech tools."