Private Markets Update & Deal Flow

Jeff Bezos is back in an executive chair for the first time since leaving Amazon in 2021, and his first public explanation of what Project Prometheus actually does landed like a puzzle. "It's a very, very modern version of CAD," he told CNBC's Andrew Ross Sorkin at the Blue Origin Rocket Park in May. Then: "I'm really oversimplifying." The company is building an "artificial general engineer" — AI tooling for designing and manufacturing real physical objects across aerospace, computing, and automobiles. Not robots. Not a chatbot. Something in between, and in some ways more interesting than either.

The round has grown since the USD 6.2B founding raise reported in late 2025. Prometheus is now said to have closed USD 10B at a USD 38B valuation, with JPMorgan, BlackRock, private equity, and sovereign wealth funds joining the cap table — a deliberate pivot away from pure Silicon Valley venture money toward institutions with manufacturing and industrial exposure. The co-CEO structure pairs Bezos with Vik Bajaj, a physicist and chemist formerly of Google X who co-founded Verily and Foresite Labs. By all accounts Bezos is hands-on in a way he wasn't at Blue Origin or Amazon in its later years.

Underneath the product ambiguity sits a bolder structural thesis. Prometheus is reportedly planning a Berkshire Hathaway-style holding structure to acquire manufacturers that would benefit from its tools — turning a software company into a physical economy play. The team of 120, drawn from OpenAI, DeepMind, Meta, and Google, gives it credibility before a single product has shipped. The first products haven't been decided yet. That's either the most concerning thing about it, or the most intriguing.

The signal: Bezos is not trying to out-Anthropic Anthropic or out-xAI xAI. He's making a bet that the most valuable AI application of the next decade is the one that collapses the distance between design intent and manufactured object. If he's right, the USD 38B entry point will look modest. If the product story doesn't materialise, the cap table will look like a very expensive job interview.

Anthropic is in advanced talks to give the European Union access to Mythos, its most restricted frontier model, launched quietly in April under Project Glasswing — Anthropic's cybersecurity initiative. Access has been limited to a tiny circle of vetted organisations for a reason: Mythos is exceptionally good at finding software vulnerabilities. In the right hands, it's a national-security tool. In the wrong ones, something considerably worse. The EU wants in precisely because it understands those risks, which is the kind of circular logic Brussels specialises in but isn't, on reflection, irrational.

The political structure underneath this negotiation is what makes it interesting. Anthropic has reportedly told the EU it requires a sign-off from Washington before sharing the model with any foreign government — placing this squarely inside the US export-control and AI-supremacy debate. The current administration has been deliberately cautious about frontier AI crossing borders; this is a test of exactly how tight those controls will be in practice. The EU already secured access to OpenAI's GPT-5.5-Cyber in May, which adds competitive pressure on Anthropic to move — and on Washington to make a decision.

The broader backdrop: Anthropic closed a USD 65B Series H at a USD 965B valuation, with run-rate revenue crossing USD 47B earlier in May. Enterprise adoption is doing the heavy lifting, though everyday Claude usage is accelerating behind it. Infrastructure commitments span up to five gigawatts with Amazon, next-generation TPU capacity with Google and Broadcom, and GPU access through SpaceX's Colossus 1 and 2. At USD 965B, the market's position is clear. The USD 1T conversation is no longer hypothetical.

Why it matters: the Mythos diplomacy is the first real test of whether frontier AI will be treated as a strategic export the way semiconductors and missile guidance systems are. The outcome sets a precedent well beyond Anthropic's cap table. Pre-IPO allocation remains open — contact us to discuss current availability.

Perplexity CEO Aravind Srinivas has a thesis that cuts against the prevailing AI narrative: the next battleground isn't who has the biggest model, it's who can deliver the most economic value per watt per user. Efficiency, not raw power, is the new moat. The company's answer is "orchestration" — an AI layer smart enough to decide which model handles which task, where computation happens, and how agents coordinate. Some of that compute, he argues, doesn't need a data center. "The data center is coming to your laptop."

The product that carries this thesis is Personal Computer — a hybrid local-cloud AI agent launched first on Mac in March, expanded to all Pro subscribers in May, and announced for Windows on June 3. The Windows version integrates with Word, Excel, PowerPoint, and Outlook, effectively making Perplexity an operating layer that sits above Microsoft's own productivity stack. At Computex, the company demonstrated hybrid inference — the system autonomously routing some work to local compute and some to the cloud, based on privacy, speed, and cost. It's still rolling out, but the architecture is the point.

The commercial result: annualised revenue has tripled this year. The bigger question is whether "best orchestrator" becomes a defensible moat, or just a gap the model giants eventually close themselves. Perplexity's counter is that it doesn't depend on any of them — when Anthropic improves, Perplexity improves too. It's a leveraged position. The risk is that OpenAI, Google, or Microsoft decides it wants that layer back.

The signal: Perplexity is building toward an IPO on the back of a structurally different argument — that the value in AI accretes to the router, not the model. The revenue trajectory says the market is buying it so far. The competitive question is whether that thesis survives the next round of model releases from the giants it depends on.

Nvidia has picked Unitree, China's leading humanoid robotics startup, as the partner for its first publicly available humanoid system built on the Isaac GR00T platform. The H2 Plus — a six-foot, 150-pound reference design combining Unitree's chassis, Sharpa's 22-degrees-of-freedom tactile hands, and Nvidia's Jetson Thor module and Blackwell GPU — was unveiled at GTC Taipei and Computex on June 1. Huang described it simply: "This platform runs the new Thor, and our entire software stack. We built this for higher education and university researchers, because for them to build this is insanely hard to do."

The institutional queue is telling: ETH Zurich, Stanford, UC San Diego, and Seattle's Ai2 are already lined up. The idea is to give labs a common hardware platform so that research can start on day one rather than after months of integration work. Nvidia confirmed it plans to repeat the exercise with humanoid makers in the US, Europe, and South Korea — Unitree is the proof of concept, not the exclusive deal. The H2 Plus goes on sale in the second half of 2026; "anyone can buy it," according to Nvidia's VP of physical AI simulation.

The timing is not coincidental. Unitree is heading for a STAR board IPO targeting roughly USD 610M, with more than 40% of revenue already coming from outside China. A Jensen co-sign on the same stage as a Computex keynote, weeks before pricing, is not a product announcement — it's a market positioning event. The catch: humanoid robotics remains early-stage, and most real deployments are still stuck in controlled warehouses. Nvidia's thesis is that seeding the research layer is how you compress that timeline. Whether it arrives before institutional patience runs out is the question the IPO will partly answer.

What Nvidia is really doing: the same thing it did to AI training. Own the developer layer before anyone else gets there, seed it into every research lab on earth, and collect on the compute as the applications mature. The US–China angle adds complexity — Huang called China "formidable" in robotics this week, and he's right. Whether the geopolitics allows this kind of collaboration to continue at scale is an open question.

The Blue Origin story has a new and considerably more sobering chapter. NASA Administrator Jared Isaacman told CNBC this week that restoring the Cape Canaveral launchpad destroyed in the May 28 New Glenn explosion is likely to take until 2028. "It will take some serious time," he said, adding that 2028 is "within the realm" of a possible recovery timeline. The damage covers the launchpad itself, the lightning tower, and the transporter erector — the kind of infrastructure that takes years to permit, procure, and rebuild, not months.

The commercial fallout is significant. Blue Origin had roughly 12 New Glenn launches penciled in for this year — that schedule is now aspirational at best. The more pressing concern sits at Amazon: the company has committed USD 2.7B to Blue Origin to help lift its Kuiper broadband constellation, and the FCC has imposed deployment deadlines that don't accommodate rocket anomalies. The Kuiper launch math now has a gap where Blue Origin's contribution was supposed to sit, which increases pressure on third-party launch providers and raises questions about schedule slippage that Amazon has publicly avoided discussing.

NASA has also written Blue Origin a USD 468M cheque for lunar lander work tied to Artemis — the agency says it will "support the investigation and assess the impact." Bezos and Blue Origin CEO Dave Limp have both been publicly optimistic about rebuilding, and Limp has noted the support tower can be repaired in place rather than demolished. The broader competitive picture is harder to spin: two missions in a row have encountered serious problems. The gap between Blue Origin and SpaceX isn't a programme gap anymore. It's a category gap.

The contrarian read remains the same: every Blue Origin stumble is a compounding event for SpaceX's moat. Starlink's position in launch, connectivity, and defence-grade communications grows stronger with each competitor setback. The SpaceX IPO window looks even more compelling when viewed against the Blue Origin timeline.