Four Rounds Are a Quarter

Crunchbase reports that global venture funding in Q1 2026 reached roughly $297 billion, of which approximately 80% went to AI. This is being written up as a record-breaking quarter, which it is, in the narrow sense that the headline number is larger than any previous headline number. It is also, if you read the underlying deal list, four companies.

What the data actually shows. OpenAI raised $122 billion. Anthropic raised $30 billion. xAI raised $20 billion. Waymo raised $16 billion. That is $188 billion across four rounds, or about 63% of the entire quarter's global venture total. Strip those four out and the remaining ~6,000 startups in the Crunchbase dataset divided roughly $109 billion between them, which is a perfectly normal venture quarter, possibly a slightly soft one depending on how you treat the long tail.

So the question worth asking is whether "Q1 2026 was a record AI venture quarter" is a useful sentence. I don't think it is. It conflates two markets that have almost nothing structurally in common.

The two markets. One market is venture capital as practised: thousands of rounds, seed through growth, priced against comparable rounds, with standard preference stacks and information rights. The other market is sovereign-scale frontier-lab financing, in which a small number of labs raise sums that exceed the GDP of medium-sized countries, in structures that look much more like project finance or structured credit than like Series F equity. Crunchbase counts both as "venture funding". They are not the same instrument.

OpenAI's $122 billion in a single quarter is the clearest case. I do not have the full term sheet, but reporting around recent OpenAI financings has consistently disclosed structured components: committed compute purchases from strategic partners, credit facilities, equity tranches with non-standard liquidity terms, and capital tied to specific data-centre build commitments. Treating the announced round size as a clean equity number, the way Crunchbase methodology appears to, materially overstates the cash-for-shares figure. There is a Vision Fund analogy here that I will resist labouring, except to note that the 2017-2019 global VC totals also looked unprecedented until they didn't.

Where the residual deal flow is going. The more interesting story is in the week-to-week list, where the non-mega rounds tell you what the rest of the market thinks is worth funding. Crunchbase's roundup for the week ending 22 May 2026 surfaces three rounds that, between them, sketch the working thesis of capital that is not chasing frontier labs:

• Nexthop AI, $500M Series B, AI networking fabric.
• Unconventional AI, $475M seed at a $4.5B valuation, neuromorphic compute, founded by ex-Databricks executive Naveen Rao, reportedly targeting $1B total mega-seed.
• Resistant AI, $25M Series B, agent-mediated fraud detection.

Networking, chips, security. Picks and shovels. Not one application-layer company in the bunch. This is consistent with the inference-economics frame: as the binding constraint moves from training to inference, and as frontier labs continue to absorb most of the available application surface area through their own products, the durable venture-scale opportunities sit one layer down. Networking fabric to move tokens between accelerators, novel substrates to compute them more cheaply, security to police the agents that consume them. The application layer is being eaten by the labs that fund themselves at $122 billion a clip, which is the structural reason application-layer rounds are getting harder to price.

Unconventional AI is the bet worth flagging. A $475 million seed at $4.5 billion, with reported plans to extend to $1 billion total, is one of the largest seed rounds on record. Rao is a credible operator: he built Nervana, sold it to Intel, then ran AI infrastructure at Databricks. The thesis is that the GPU substrate is wrong for inference-dominated workloads, and that a neuromorphic architecture can deliver the efficiency gains that GPU roadmaps cannot.

I want to be careful here. Neuromorphic compute has been "the next paradigm" for about a decade, and the graveyard of non-Nvidia accelerator companies is well-populated: Graphcore, Cerebras, Groq in its earlier form, several others. The track record is not encouraging. But the round is not really a bet on neuromorphic per se; it is a bet that inference economics will degrade GPU margins faster than Nvidia's roadmap can compensate, and that when that happens, a credible alternative substrate will be valuable enough that getting it to silicon now is worth $1 billion of seed capital. This is the kind of bet that is either extremely well-timed or extremely wrong, with very little in between.

It is also, in the intelligence-explosion-signals frame, the clearest venture-scale wager that compute-efficiency breakthroughs are imminent enough to underwrite. I'd watch for two things: who follows on at Series A (strategic vs. financial composition will tell you whether the existing chip incumbents see the thesis as a threat), and whether Unconventional ships first-silicon performance numbers within 18 months. If they do, the inference-economics frame sharpens considerably. If they don't, this becomes another expensive chip-company cautionary tale.

The Polymarket adjacency. Polymarket has separately announced prediction markets on Anthropic and OpenAI private-company share prices. This is a small thing structurally but a clean signal: secondary-market liquidity infrastructure is being built ahead of any IPO, which implies the market thinks pre-IPO price discovery has enough volume to be tradeable. Whatever the implied valuations settle at, the existence of the contract tells you that the private rounds are now being treated as price-discovery events for an eventual public listing, rather than as terminal capital events. This is not how Series F rounds usually function. It is how late-stage structured financings function before an IPO.

What this is a case of. It is a case of the venture asset class bifurcating, with the top decile of rounds now sitting in a different financial instrument from the rest of the asset class while sharing a label. The "80% AI" figure is real, but it is doing work the underlying reality doesn't support. If you are a non-AI founder reading the Q1 number and concluding that capital has fled your sector, the more accurate read is that capital has not fled — it has been numerically swamped by four rounds that are not really comparable to yours.

What to watch.

• Whether Crunchbase or KPMG eventually break out equity-only versus total-round-size figures for the frontier-lab megarounds. The methodology question is the whole story.
• Unconventional AI's Series A composition and silicon timeline.
• Whether Q2 2026 produces a second wave of infrastructure rounds in the $200M-$500M range, or whether Nexthop and Unconventional were the cluster.
• Polymarket's volumes on the OpenAI and Anthropic contracts. If they trade thinly, the IPO signal is weak; if they trade with depth, the private-market is now functionally a public one.

Footnotes

¶