Analysis
Two things happened this week that look unrelated until you put them side by side. Nvidia fell 3.39% on Monday, extending a slide that has now wiped out roughly $1 trillion in market value since the stock's record high on May 14 -- a drawdown severe enough that NVDA now trades around 18 times forward earnings, below the S&P 500's roughly 21x multiple and the cheapest the stock has traded since early 2019. In the same 48 hours, Meta confirmed its Louisiana "Hyperion" data center will cost more than $50 billion, and Munich-based Helsing closed a $1.8 billion Series E at an $18 billion valuation, Europe's biggest-ever defense-tech round, raised explicitly to build more AI-driven autonomous systems. Compute demand isn't slowing down. It's the compute seller who's getting repriced.
Start with Nvidia, because the drivers matter more than the headline number. Hawkish Fed signals have raised cost-of-capital concerns across high-growth tech broadly, which hits a stock trading on a growth multiple harder than most. But the more AI-specific driver is reporting that major hyperscale cloud customers -- the exact companies whose capex is supposedly the bull case for Nvidia -- are accelerating a shift toward custom ASIC chips designed to reduce reliance on third-party hardware. Layer on gross-margin pressure from higher memory costs, increasingly crowded institutional ownership, and criticism that Nvidia is using its cash pile for vendor financing of its own customers rather than buybacks or dividends, and you get a stock that's fallen out of favor even as the underlying AI buildout keeps accelerating.
That buildout is exactly what Meta's Louisiana numbers show. Hyperion, Meta's largest data center, is now a 5-gigawatt facility costing more than $50 billion -- up from the $27 billion figure Meta and Blue Owl Capital disclosed when they formed their joint venture in October 2025, which was itself already an increase from an original plan of roughly $10 billion and 2 gigawatts. That's not a modest cost overrun; it's the estimate roughly doubling again in well under two years, even as Louisiana's governor signed a 20-year sales-tax exemption for data centers built before 2029 and local businesses have already collected more than $1.6 billion in Meta contracts since construction began in December 2024.
“Start with Nvidia, because the drivers matter more than the headline number.”
Helsing's raise is the same demand signal from a completely different sector. Europe's answer to Anduril pulled in $1.8 billion at an $18 billion valuation, with Goldman Sachs Alternatives, Dragoneer, Iconiq, CPPIB and JPMorgan joining existing backers Lightspeed, General Catalyst and Plural -- and investor demand reportedly exceeded the available allocation by a wide margin. It lands alongside a US defense-tech pace that's just as aggressive: Anduril's $5 billion Series H at a $61 billion valuation in May, Shield AI's $2 billion Series G, Saronic's $1.75 billion Series D, Mach Industries' $300 million round at a $1.8 billion valuation. None of that capital is being deployed to buy merchant GPUs at list price -- it's building proprietary, vertically integrated AI systems.
Put the numbers next to each other and the pattern is stark: roughly $1 trillion erased from the largest AI chip seller's market cap, against a Louisiana data center whose cost estimate has 5x'd from its original plan and a European defense-AI round that came in oversubscribed. Meta's own in-house chip effort -- the MTIA-based "Iris" chip built with Broadcom and manufactured by TSMC, set to enter production in September and help scale Meta's compute from 7 gigawatts this year to 14 next year -- is a concrete example of exactly the dynamic Nvidia's selloff is pricing in: the biggest buyers increasingly want to own the chip layer themselves, not rent it from a single merchant supplier.
For VCs and founders building anywhere near AI infrastructure, the read-through is specific: capital and margin are migrating toward whoever controls the customer relationship, the site, or the proprietary system -- hyperscalers building their own silicon, defense primes building their own autonomous stacks, data-center operators locking in multi-year power and land commitments -- and away from being a fungible, if currently dominant, component supplier. That's a very different portfolio-construction question than "how much AI capex is there," which remains unambiguously high.
The bear case for this thesis is that Nvidia isn't going anywhere -- its CUDA software moat, its multi-year lead in training-cluster performance, and its sheer scale advantage over any single hyperscaler's ASIC program are real, and a few hyperscalers diversifying their chip mix at the margin doesn't mean Nvidia's core franchise is broken. A single week's stock move driven partly by macro rate fears shouldn't be overread as a structural thesis on its own. What to watch next: whether Nvidia's next earnings call shows any actual revenue impact from hyperscaler ASIC adoption rather than just investor anxiety about it, and whether Meta's own Iris chip ships on schedule in September at the volumes needed to meaningfully dent its Nvidia purchases.
