AI Investment Since 10 April: Deal Tracker and Analysis

01Scale of capital over six months

Private valuations at frontier labs now approach the trillion-dollar mark, public listing windows have opened simultaneously in New York and Shanghai, chipmakers are aggressively acquiring software and model ecosystems up the stack, and China’s leading self-funded model lab took in hundred-billion-yuan external capital for the first time.

02Deal timeline (10 April – 22 October)

03Capital concentrates at the top while the second tier consolidates

The largest checks of the half-year flowed into four or five frontier labs: Anthropic raised $65B in a single round ($965B post-money), OpenAI is negotiating $30B above a trillion-dollar valuation, DeepSeek raised two massive rounds taking it from ¥340B to ¥500B in six months, and Moonshot AI climbed from $20B to $50B in five months.

At this scale, private rounds function as heavy infrastructure project finance — almost every dollar goes straight into GPUs, data centers, and power contracts dictated by the compute budget of the next frontier training run.

Squeezed by trillion-dollar giants, second-tier labs are consolidating. Cohere and Aleph Alpha signed a ~$20B merger, while Mistral raised €3B in Series D funding at a €21B+ valuation.

04IPO windows open simultaneously in the US and China

Six months ago, M&A was virtually the only exit route for AI startups. Today, listing pipelines are active across both US and Chinese exchanges:

Market dynamics differ sharply between the two. In New York, SpaceX raised $75B and gained 19% on day one as institutional funds priced a trillion-dollar asset. In Shanghai, Unitree surged 629% at the open to ¥445B, driven by intense domestic retail demand for robotics and hard tech.

Neocloud provider Nscale will be an important litmus test. Its revenue is heavily concentrated in Microsoft and Anthropic, and less than a month after filing, Microsoft terminated a major UK lease over power delays. How public markets price customer concentration and power risk will set the bar forLambda and other compute providers in the queue.

05Chipmakers buy their way up the software and model stack

The dominant M&A pattern over the past six months is silicon vendors acquiring software ecosystems and model labs: Nvidia agreed to buy Hugging Face for $12.9B, locked in Poolside for $7B ($6B licensing + $1B equity), entered talks to buy Reflection for ~$25B, and backed inference chipmaker d-Matrix; AMD acquired World Labs for $8.2B; Qualcomm bought Modular and is in talks for Tenstorrent.

The strategic motive is clear: whichever developer hub, compiler stack, or model architecture wins adoption dictates which hardware gets bought. Buying developer platforms and frontier model teams locks in downstream silicon demand.

Meanwhile, incumbent enterprise software vendors are aggressively buying vertical AI applications: Salesforce acquired Fin, Listen Labs, and Braze; Autodesk bought MaintainX; Stripe acquired model router OpenRouter. The largest application-layer deal was SpaceX’s $60B all-stock acquisition of coding agent Cursor, while rival Cognition saw its valuation jump from $26B to $40B+ in under three months.

06Circular vendor financing across silicon, clouds, and labs

Placing Google and Amazon’s milestone investments in Anthropic alongside Anthropic’s $35B compute commitment to Lambda and Nvidia’s off-balance-sheet guarantees reveals a tight capital loop across the supply chain:

Vendor financing is standard practice during capital-intensive infrastructure booms — the late-1990s telecom fiber build-out used the exact same playbook. It accelerates physical deployment while making aggregate demand look larger than organic end-user revenue alone.

Evaluating headline deal numbers requires stripping out circular credits: how much of a “$10B round” is unrestricted cash versus cloud credits that flow right back to the investor? Nvidia’s quarterly disclosure of up to $164.5B in off-balance-sheet lease and supply guarantees (including backing OpenAI’s Ohio campus lease) puts a concrete figure on the leverage embedded in the build-out.

07Grid bottlenecks replace chip shortages as top-end valuations soften

Electrical grid capacity has overtaken GPU supply as the binding constraint on cluster expansion. In April, OpenAI paused Stargate UK over power costs; in October, Microsoft walked away from a 23,000-GPU Nscale lease in the UK because the local utility could not connect power until the 2030s.

At the same time, late-stage valuations are showing cracks: OpenAI’s pre-money valuation stepped down from $1.4T in September talks to $1.25T in October, and Thinking Machines Lab priced below its initial $50B+ target. Large rounds increasingly rely on tranched, two-step structures — granting a lofty headline valuation on paper while gating actual cash payouts behind steep operational milestones.

At the seed stage, however, pricing remains frothy: Emulate, a one-month-old world-model team with no product, is negotiating a seed round at a $3–3.7B pre-money valuation.

08Four takeaways for early-stage investing

1. 01Both IPO and M&A exit channels are wide openWith public listings and multi-billion-dollar acquisitions clearing simultaneously, early-stage funds have tangible liquidity paths again.
2. 02Follow the incumbents’ shopping listsChipmakers are paying up for compilers and developer hubs; legacy SaaS giants are buying vertical AI workflows. Those exact layers offer the clearest M&A premiums.
3. 03Ignore frontier lab valuations when pricing startupsTens-of-billions rounds at frontier labs are infrastructure project finance. Using them as comparables for early-stage software companies guarantees overpaying.
4. 04Dissect deal structure behind headline valuationsStrip out compute credits, milestone tranches, and structured liquidation terms before taking any headline valuation at face value.