Fewer bets, bigger checks: the seed drought hiding inside a $17 billion AI week

The headline number for the week of May 4 is $17 billion. That sounds like momentum. It is, partly. But the composition of that number tells a different story from the aggregate, and the composition is where the actual signal lives.

Seed-stage deal count fell 28% week-over-week, from 32 rounds to 23. Series C and D counts went the other direction, up 133%, from three rounds to seven. Median check size jumped from $15M to $19.5M. Total capital deployed rose 75% while total round count fell 8%. The market is not accelerating broadly. It is concentrating: fewer bets, higher conviction, much larger checks at the later stages. Early-stage founders competing for seed capital this week were operating in a tighter market than the $17B aggregate figure suggests.

The most structurally interesting stories of the week were not the biggest dollar amounts. SAP acquired two AI companies in two days to solve a problem that the entire LLM industry has been quietly ignoring. China's Moonshot AI raised $2B for an open-source model that is already second in usage on OpenRouter. And four European quantum hardware companies raised $430M in one week, an amount that is hard to explain without a thesis.

The numbers

Three rounds account for $6B of the $17B total: Isomorphic Labs ($2B, Alphabet-backed drug discovery AI), Moonshot AI ($2B, Chinese open-source LLMs), and Esentia Energy ($2B, energy infrastructure with no AI component in our database). Strip those three and the remaining 83 rounds represent about $11B. Still a strong week, but the mega-round effect is real and should be separated from the trend line.

Seed is draining while late-stage concentrates

The 28% drop in seed rounds over a single week could be noise. But the direction matches a pattern visible across the trailing four weeks in our database: the share of total rounds classified as seed has been drifting down, while Series C and D rounds, historically rare in any given week, keep appearing. This week produced seven late-stage rounds with disclosed amounts averaging well above $100M each.

The clearest interpretation is that institutional capital has decided which bets it wants to press. Sierra raised $950M at a $15.8B valuation, led by Tiger Global and GV, according to reporting from SiliconAngle and Axios. The company crossed $150M in ARR in February 2026, serves 40% of the Fortune 50, and its growth rate made the $950M round look, by the logic of revenue multiples, almost conservative. Ramp raised $750M in a round that reportedly values it near $40B, six months after its last $32B round. These are not seed-stage bets made on founders and theses. They are growth-stage commitments to companies with measurable revenue and enterprise contracts.

The implication for founders at the pre-seed and seed level is uncomfortable. Capital is not scarce in aggregate, but the flight-to-quality dynamic means that investors capable of writing $10M-$30M seed checks are increasingly being asked by their LPs to explain why they are not deploying into known winners at growth. The result is a bifurcated market where a well-timed growth round has never been easier to close and a first institutional seed round has not gotten meaningfully easier since 2023.

Open-source AI at $20B: the Moonshot AI signal from China

Moonshot AI closed a $2B Series D at a $20B+ valuation, led by Meituan's venture arm. Six months ago, the company was valued at $4.3B. The 4.6x increase in six months is the fastest valuation expansion among Chinese AI labs this cycle, according to Huafeng Capital data cited in TechCrunch.

The product behind this is Kimi K2.6, an open-source model released April 20 under a Modified MIT license. As of the funding announcement, Kimi K2.6 was the second-most used LLM on OpenRouter, behind only DeepSeek. Moonshot's ARR crossed $200M in April 2026. These are not theoretical numbers: this is a Chinese open-source model that reached $200M ARR faster than most Western open-source AI companies and is being used more than every OpenAI and Anthropic model on a major distribution platform.

The structural point here matters more than the individual company. When an open-source Chinese LLM reaches $200M ARR and a $20B valuation in the same month, it changes the calculus for Western foundation model labs. The argument that closed, proprietary models would maintain premium pricing because of capability advantages is harder to sustain when open-weight models with comparable benchmarks are freely available. DeepInfra's $107M Series B this same week, from a company processing nearly five trillion tokens per week across 190+ open-source models, is the demand-side confirmation: inference spend on open-source is no longer speculative infrastructure. It is production workload.

SAP bought tabular AI: what that tells you about the LLM stack

On May 4 and 5, SAP announced two acquisitions in 48 hours: Prior Labs for $1.16B and Dremio for an undisclosed sum. The pairing is precise and deliberate in a way that deserves attention.

Prior Labs builds tabular foundation models (TFMs), a category of model trained specifically on structured data like spreadsheets, SQL tables, and ERP records rather than text. Large language models are notoriously poor at tasks that require precise numerical reasoning over structured rows and columns. SAP's own press release noted this directly: "Large language models are focused on text and don't operate effectively on tables, numbers, and statistics. TFMs do." SAP is committing to invest more than one billion euros over four years to scale Prior Labs into a frontier AI lab for structured data.

Dremio is the data access layer: an open, high-performance data lakehouse platform that lets SAP's AI models query both SAP and non-SAP data in real time without traditional ETL pipelines. The combination is a deliberate attempt to make SAP's AI stack self-sufficient: the TFM from Prior Labs runs on the data fabric from Dremio, without sending data to OpenAI or Anthropic.

This is the clearest indicator yet of what enterprise software incumbents actually intend to do with AI. They are not building chatbot wrappers on top of GPT-4. They are acquiring the specific AI capability their core product is missing (structured-data reasoning), acquiring the infrastructure to run it on their customers' data, and explicitly framing it as a European AI sovereignty play. SAP CEO Christian Klein said the Prior Labs acquisition would "establish a globally leading frontier AI lab in Europe." The two deals together cost SAP north of $1.16B plus whatever Dremio's acquisition price was, and signal a category that will see more M&A: companies that build foundation models specialized for structured, tabular enterprise data.

Four quantum rounds in five days: real inflection or sector clustering?

Quantum computing raised more capital in the week of May 4 than in many full months from 2022. Four rounds closed or were announced: QuantWare (Netherlands) raised a $164M-equivalent Series B to build quantum processors at industrial scale, the largest private round ever raised by a dedicated quantum processor company according to EU-Startups. Quantum Motion (UK) raised $160M Series C for silicon-based quantum computing using CMOS transistor technology. eleQtron (Germany) raised a 67M euro Series A for trapped-ion quantum systems. QuTwo raised a $27M angel round for quantum AI, per Tech.eu.

The geography is striking: Netherlands, UK, Germany, and a fourth European company in one week. This is not coincidental. Several European nations have active government quantum programs, and the defense-and-sovereignty context created by geopolitical shifts has accelerated institutional willingness to fund deep-tech hardware. Intel Capital joined the QuantWare round, which signals US strategic interest in European quantum hardware rather than pure financial investment.

Whether this represents a real quantum computing inflection is genuinely unclear. The companies are at different technical approaches (superconducting qubits for QuantWare, silicon CMOS for Quantum Motion, trapped ion for eleQtron) and none have announced commercial deployments at meaningful scale. What is clear is that hardware-level quantum investment in Europe has reached a pace that was not present 18 months ago. The AI-specific angle is thinner than the funding announcements imply: most of the value proposition is quantum advantage for chemistry, materials science, and logistics optimization rather than LLM-adjacent workloads. Treat these as deep-tech hardware bets with long time horizons, not the next rung of the AI infrastructure stack.

Voice AI becomes institutional

ElevenLabs' third close of its Series D brought participation from BlackRock, Wellington Management, NVIDIA's NVentures, Santander, and D.E. Shaw alongside Jamie Foxx, Eva Longoria, and the creator of Squid Game. The round total is now above $550M. ARR crossed $500M in the first four months of 2026, up from $350M at the end of 2025.

The participation of BlackRock and Wellington is the structural signal. These are not venture firms that take positions for optionality. They are institutional investors that move into companies they expect to hold equity in for years, often as pre-IPO positioning. ElevenLabs at $11B valuation and $500M ARR is priced at roughly 22x revenue, which is high but not irrational for a company growing at this pace in a market where voice AI is quickly becoming a required enterprise capability rather than a novelty.

The celebrity investors (Foxx, Longoria, McConaughey as an existing investor) and the Squid Game creator tell a different part of the story: the entertainment industry is putting money into the infrastructure that will eventually be used to clone, license, or generate voice at scale. That is not a comfortable sentence, but it is the economic logic behind A-list entertainment figures taking Series D positions in a voice synthesis company.

Microtrends worth watching

• Insurance found its AI-native moment. Corgi raised a $160M Series B at a $1.3B valuation, four months after its $108M Series A. That is a 10x valuation increase in four months, which is either exceptional product-market fit or a market temporarily overreacting to AI-native insurance framing. Reserv raised a $125M Series C from KKR to accelerate AI-driven insurance claims processing. Three distinct AI insurance plays closed rounds this week. The category is real; whether three unicorns can coexist in startup-focused insurance is the open question.
• Physical AI language drift in robotics. New startups entering the database this week used terms that did not appear in bulk 12 months ago: "Physical AI," "Embodied AI," "Vision-Language-Action Model," "Large Action Model," and "Robot Foundation Model." Multiple companies, including Persona AI and Manifold AI, explicitly described their products as foundation models for physical systems rather than robots with AI features. This is a meaningful reframing: it positions humanoid robotics in the same language category as LLMs and makes the fundraising pitch structurally similar.
• Agent identity as an emerging infrastructure category. New startups this week included at least one explicitly building "The Agent Identity Standard for autonomous systems," describing itself as authentication and credentialing infrastructure for AI agents acting autonomously on the internet. As agentic AI systems proliferate, the question of how to verify, authorize, and audit AI agents in multi-agent pipelines is genuinely unsolved. This is a small signal, but it is the kind of infrastructure category that looks obvious in retrospect.
• Autonomous finance is quietly clustering. Fazeshift raised $17M Series A, explicitly describing its product as "autonomous finance" with AI agents executing financial operations end-to-end. Combined with Ramp's $750M round, several smaller fintech AI rounds, and Nace.AI's $21.5M seed for a "metamodel" that autonomously runs enterprise workflows, the pattern is clear: financial operations automation is the vertical where agentic AI is finding the easiest product-market fit, because the tasks are well-defined, the data is structured, and the cost of error is measurable.
• SAP and Databricks each absorbed an AI company this week. Agent Bricks was acquired by Databricks. Prior Labs and Dremio were acquired by SAP. Two AI agent infrastructure companies got absorbed by enterprise platforms in five days. The M&A window for early-stage AI infrastructure is open.

What might happen next week

Two predictions, both specific enough to be wrong:

First, the quantum hardware funding surge will attract a counternarrative. At least one skeptical analysis piece will publish questioning whether the $430M raised by quantum hardware companies this week reflects genuine technical progress or is primarily defense-adjacent geopolitical spending dressed up as commercial enterprise. The skeptics will have a point worth engaging: none of the companies funded this week have announced commercial customers at production scale, and the gap between the funding velocity and the commercial deployment reality in quantum computing has historically been large. Watch for the backlash and whether any of the funded companies respond with deployment metrics.

Second, there will be another "tabular AI" or "structured data AI" acquisition announcement within three weeks. SAP's Prior Labs deal surfaced a small number of comparable companies building TFMs and structured-data reasoning layers, and at least one other enterprise software company (the obvious candidates are Salesforce, Oracle, or ServiceNow) has likely been watching the same category. The Prior Labs deal priced the category at roughly $1B+ for a company that had not yet shipped at scale. That price signal is now in the market.