Baseten tripled to $13B while nobody was watching: inference infrastructure and sovereign AI defined this week's $7.3B in deals

Baseten does not make AI models. It runs them. The company's software optimizes inference for open-source models, handling the latency, GPU scheduling, and cost-per-token arithmetic that determines whether an AI application is profitable or a money pit. On June 18, it closed a $1.5 billion funding round at a $13 billion valuation. Five months ago, Baseten was worth $5 billion. Its annualized revenue grew from $200 million to $600 million between December 2025 and March 2026, a reported 1,900% year-over-year increase. Altimeter Capital, Conviction, Spark Capital, Sands Capital, and Wellington Management co-led.

That trajectory makes a specific argument about where value is concentrating in the AI stack. The model builders are large and visible. The inference layer is quieter, apparently faster-growing, and now worth more per dollar of revenue than almost any model company not named OpenAI or Anthropic. Baseten's customers (Cursor, Mercor, OpenEvidence) report 30% cost savings versus closed-source APIs. When you cut 30% from the input that determines your unit economics, that vendor becomes load-bearing infrastructure.

Salesforce's $3.6 billion acquisition of Fin (formerly Intercom's AI customer service platform, announced June 15) lands the same message from a different angle. Salesforce did not buy Fin for its chatbot. It bought a production-grade AI orchestration layer to plug into Agentforce. Fin resolves 76% of support volume end-to-end. What Salesforce paid for was the proven runtime, not the underlying model. Two of the week's biggest deals were, at their core, bets on execution infrastructure rather than model capability.

The numbers

Total disclosed capital this week: $7.3 billion across 61 rounds. That is down sharply from $44.7 billion last week, but the comparison is almost entirely a function of three anomalous rounds from June 8 to 14: Prometheus ($12B), Helix Digital Infrastructure ($10B), and Supermicro's equity deal ($7B). Strip out every round above $1 billion from both periods and the picture stabilizes into a more moderate decline.

The underlying pattern at the sub-billion level shows a moderate 34% decline. The AI sector's share of disclosed capital rose 7 percentage points this week despite the overall total falling, suggesting non-AI sectors contracted more sharply than AI did.

Inference infrastructure absorbed $2.2 billion and most people called it a quiet week

Baseten leads a clear cluster: Baseten ($1.5B, model serving infrastructure), DayOne Data Centers ($600M, GPU infrastructure in Asia backed by Hillhouse Capital), HyperLight ($80M, photonic AI accelerators), and Aston Power ($20M, power infrastructure specifically targeting AI data center demand). Four companies in the physical and software stack beneath the model layer raised $2.2 billion in one week.

The arc that connects them: open-source models are commoditizing faster than most forecasters predicted two years ago. That commoditization makes inference more valuable, not less. Every new AI application that launches needs to run inference on something. Baseten's competitive claim is that it runs open-source models cheaper and faster than closed-source APIs, with customers reporting 30% cost savings. At $600M ARR in March, growing nearly 10x year-over-year, it is the canonical example of a company whose growth is driven entirely by the success of other companies it never owns. The more AI startups launch, the more customers Baseten acquires.

The $13 billion valuation will strike some observers as aggressive for an infrastructure business. The counter-argument: Baseten is not a services business, it is a leverage business. Each customer that routes inference through Baseten instead of a closed-source API increases throughput and revenue without a proportional increase in cost. If the trend of model cost compression continues, the economic case for open-source inference gets stronger each quarter. The risk is obvious: AWS, Azure, or Google builds a competing product and prices it at marginal cost to protect their cloud revenue. So far, they have not.

Sovereign AI hardened from a talking point into a funding category

Dream, the Israeli sovereign AI and cyber defense company, raised $260 million on June 18 at a $3 billion valuation. Co-led by Bicycle Capital and Group 11, with Bain Capital Ventures and Tru Arrow Partners participating. Founded in 2023 by Shalev Hulio, former Austrian Chancellor Sebastian Kurz, and Gil Dolev, Dream serves governments and critical infrastructure operators across Europe, the Middle East, and Southeast Asia. Its annual revenue from six government customers is $130 million. The company's framing is deliberate: it calls itself a sovereign AI company, meaning AI infrastructure that a government can control, audit, and operate without dependence on a foreign foundation model provider.

Three days earlier, Sarvam AI closed the first tranche of its Series B at $234 million with a $1.5 billion post-money valuation, the highest reported Series B valuation in Indian startup history. The unusual detail: HCLTech led with $150 million from its own balance sheet, not through a fund. That is a strategic bet, not financial capital allocation. Sarvam is India's full-stack sovereign AI play: its speech models transcribe over 500,000 hours of audio monthly across Indian languages, and its vision model is digitizing 35 million pages of insurance forms and legacy land records. Bessemer Venture Partners, Khosla Ventures, and Peak XV Partners also participated.

Two sovereign AI rounds in the same week, from two separate regions, targeting two different sets of governments, both closing in the same 72-hour window. Nations watching each other's AI capabilities are accelerating parallel investments. The logic is consistent across both companies: if AI infrastructure is as consequential as governments now believe, having a foreign country control that infrastructure is a strategic risk. "Sovereign AI" is moving from a policy speech phrase to a funded reality with specific product categories, go-to-market motions, and growing revenue.

Kling AI's $2B round is a geopolitical signal as much as a fundraise

Kling AI, the video generation unit being spun out of Chinese short-video platform Kuaishou, is seeking $2 billion at an $18 billion valuation, with General Atlantic in discussions to lead. The deal is not closed. Kling's metrics justify attention: $300 million in annualized revenue in January 2026, growing to $500 million by April following the launch of Kling 3.0. Full-year 2025 revenue was $153 million. The company plans to IPO in 2027.

The signal is not the video generation market, which everyone knows is competitive. The signal is that General Atlantic, a major American private equity firm, is apparently prepared to lead a $2 billion round into a Chinese AI company after more than 18 months of heightened restrictions on US-China technology capital flows. If this deal closes, it would mark the most significant US-China AI capital bridge in at least two years. Kuaishou's 700-million-plus user base and commercial flywheel give Kling distribution that most AI video competitors do not have. That distribution, more than the model quality, is probably what justifies the $18-to-$20 billion valuation range.

Science AI crossed a new threshold

CuspAI, the Cambridge-based materials discovery startup, is reportedly raising $400 million at a $2.6 billion valuation, with Jeff Bezos' Bezos Expeditions and Kleiner Perkins among investors. Its advisory board includes Geoffrey Hinton and Yann LeCun. The platform generates material designs and validates them in simulation; customers include ASML, Meta, and Hyundai. The $2.6 billion valuation represents a more-than-fivefold increase from its $520 million valuation in September 2025, itself following a $30 million seed in June 2024. The pace of valuation compression into this single company is unusually fast for a science-applications startup.

Radical Numerics raised $50 million in seed funding this week. Its description uses the phrase "General Biological Intelligence," defined as AI that operates on biological systems at a fundamental level, not as a drug target search tool but as an agent in biological simulation. The term is ambitious and maps loosely to what foundation models did for text, applied to biology. Whether that aspiration resolves into a product is the open question, but a $50 million institutional seed check suggests LPs find the bet worth making at this stage.

Odyssey raised $310 million in its Series B at a $1.45 billion valuation, led by Natural Capital, with Amazon, AMD Ventures, GV, EQT, and IQT participating. Odyssey builds world models: AI systems that simulate physical reality rather than predict text. Founded by self-driving vehicle veterans Oliver Cameron and Jeff Hawke, the company gathered real-world video data to build a simulation layer for training autonomous systems. AWS is the preferred cloud provider; Trainium chips are in the stack. World models for physical simulation are, in the long arc, foundational for robotics, autonomous vehicles, and synthetic data generation. The Series B is a bet that this capability is ready to be productized.

What Salesforce's $3.6B Fin acquisition signals for the rest of the CRM market

Salesforce acquired Fin (formerly Intercom's AI customer service platform) for $3.6 billion to plug production-grade AI orchestration into Agentforce. Fin resolves 76% of support volume end-to-end across live chat, email, WhatsApp, SMS, phone, and Slack. The deal closes in Salesforce's fiscal Q4 2027 pending regulatory approval.

The downstream pressure on the rest of the market is larger than the deal itself. Microsoft (Dynamics), HubSpot, and Zendesk now face the same calculus: build or buy a proven AI orchestration layer before their customers start migrating to something that handles 76% of support volume automatically. Fin's $3.6 billion acquisition price establishes a market comparable for what proven AI customer service orchestration is worth. For Series B AI customer service startups, this is a pricing data point in a future acquisition conversation. Respond.io's $62.5 million Series B, announced the same week, suggests the category still has room for multiple players before the next consolidation wave.

Microtrends worth watching

• GEO audits are emerging as a startup category. Findingyou, one of this week's newly published startups, describes its offering as "GEO audits for AI search engine citation," tracking whether brands appear in ChatGPT responses, Google AI Overviews, and MCP server discoverability. The term "Generative Engine Optimization" appeared in multiple startup descriptions added to our database this week. This is SEO for the agentic layer, and dedicated companies are beginning to organize around it.
• Voice AI is quietly stacking rounds. Bland AI ($50 million) and Respond.io ($62.5 million) both closed rounds this week with minimal press attention. Bland AI builds for AI phone agents, targeting the inbound and outbound call replacement market that enterprises want but have been slow to trust. Two rounds totaling $112.5 million in one week for the voice and conversational AI layer is not incidental.
• Defense AI is forming its own funding tier outside the US. Comand AI ($32 million, European defense AI), HighGround ($6.5 million, defense intelligence platform), and Dream ($260 million for governments) all closed explicitly B2G defense-AI rounds this week. None are based in the United States. The US defense AI flow appears to be running through larger undisclosed channels, while European and Israeli startups are operating openly in this category.
• Atom Computing's $300M quantum round is adjacent to AI infrastructure. Quantum computing has been a long-cycle bet. Atom Computing's Series C is notable because the stated applications overlap with AI-relevant simulation problems, directly adjacent to CuspAI's materials discovery thesis. If quantum begins solving problems that classical inference cannot, the data center infrastructure thesis changes in ways that are hard to model today.
• Behavox at $175 million Series D. Behavox analyzes financial firm communications for regulatory violations. $175 million at a mature stage suggests financial services AI compliance is still growing faster than the market expects. This category was considered solved three years ago. It apparently was not.

What might happen next week

Baseten at $13 billion and $600 million ARR growing at 1,900% year-over-year is past the revenue threshold many cloud infrastructure companies have used to justify S-1 filings. An IPO announcement from Baseten, or from a comparable inference infrastructure company, is plausible within 12 months. If the Kling AI General Atlantic deal closes, the IPO planning timeline should compress further for both companies. Watch for formal IPO signals from the inference layer in Q3 2026.

Sovereign AI will attract a dedicated fund announcement before the end of 2026. Two rounds above $200 million in one week from geographically distinct companies targeting different government customers is the kind of category-definition event that causes LPs to request sovereign AI exposure. A "Government AI" or "National AI Infrastructure" fund from a major VC or PE firm is the likely next move. Dream and Sarvam established pricing anchors: government AI customers pay for contracts measured in tens of millions, and the market is large enough to support multiple regional winners simultaneously.