Meta shares fell 4.3% at Thursday's open after Reuters reported the contents of an internal memo laying out the next phase of the company's AI infrastructure program.
The stock has clawed back part of the loss through the morning but stayed solidly red while the tape digested the same question it has been chewing on for nine days: is Meta the hyperscaler that just started exercising capex discipline, or the one that just committed to doubling?
Three things to note from today's news. The first is silicon. Iris, Meta's in-house AI accelerator and one of four planned MTIA generations unveiled in March, enters production at TSMC in September after clearing bug validation in six weeks with no major issues - an unusually clean result for a program that has stumbled for more than half a decade. Broadcom is the design partner under an agreement extended through 2029, and Meta plans to ship a new chip roughly every six months through 2027, against an industry norm of annual-or-slower cadences. The chips are meant to augment, not replace, externally sourced GPUs - Meta separately holds a multiyear agreement with AMD covering up to six gigawatts of Instinct accelerators - but the internal memo is very blunt about why the program matters - as adopting the latest external GPUs at Meta's scale "has been a heavy lift, and it has cost us time."
The second is scale. Meta plans to deploy seven gigawatts of computing infrastructure this year and to double overall capacity to fourteen gigawatts in 2027, with 2026 spending running as high as $145 billion - the very top of the range guided in April, and a meaningful slice of the more than $700 billion Big Tech is projected to pour into AI this year.
The third is supply. The memo reveals long-term contracts for memory from Samsung, flash storage from Sandisk and fiber-optic equipment from Sumitomo Electric - multi-year lock-ins struck in the middle of a memory shortage severe enough to be raising consumer hardware prices.
On its face the chip news is bullish: faster, cheaper, more independent compute is exactly what a company spending $145 billion a year should want. But the market has spent the past week and a half developing a very specific allergy, and the memo triggered it.
When Bloomberg reported at the start of the month that Meta was standing up a cloud business - internally, Meta Compute - to sell surplus capacity and token-metered API access to outsiders, the stock ripped nearly 9% higher in a session while CoreWeave and Nebius fell double digits. We suggested this might be a potential first crack in the AI capex boom: hoarding compute stops making sense the moment you admit you have extra, and if management appears willing to monetize idle infrastructure, the market reads capital discipline and pays for it. Days later, leaked town-hall remarks in which Zuckerberg conceded that agent development "hasn't accelerated in the way we expected" knocked the stock back down - the July 2 drop that Thursday's open just eclipsed.
Against that backdrop, a memo describing a doubling of capacity, a six-month silicon cadence and years of locked-in component supply looks rather - undisciplined when it comes to capex. Companies do not sign multi-year memory contracts in the middle of a shortage in order to stand still. As we noted earlier this month - the pivot to rewarding CapEx cutters - has, for now, been a driving force: up on plans to sell capacity, down on plans to double it, with the same infrastructure underneath both headlines.
What Doubling Actually CostsHere's the math. Fourteen gigawatts against seven implies roughly seven incremental gigawatts next year. The long-standing rule of thumb - which Jensen Huang himself used last fall around the Nvidia-OpenAI deal - is that a one-gigawatt AI data center runs about $50-60 billion of capex, roughly $35 billion of it Nvidia GPUs, and implies on the order of half a million chips drawing as much power as 750,000 homes.
It was $50 billion in September 2025. How exactly is anyone meant to afford this at this point? https://t.co/N8skNRNP9I pic.twitter.com/fng1hb21eo
— Ed Zitron (@edzitron) July 5, 2026On that math, Meta's incremental seven gigawatts carry a bill in the $350-400 billion range.
Some context:
1GW energy = enough power for 750,000 homes
1GW data center = $50-$60BN in capex spend
1GW = 500,000 GPUs
Keep in mind that the per-gigawatt price is moving up, not down. Vera Rubin - Nvidia's next data-center GPU platform - draws far more power per chip than the generation before it, so a gigawatt of capacity now holds fewer GPUs, each more expensive than the last. That is why Huang has lately floated build costs of up to $100 billion per gigawatt - a price at which Meta's seven incremental gigawatts would run roughly $700 billion. Treat the top-end number with suspicion - Jim Chanos has argued since last September that Nvidia's per-gigawatt math sits well above what operators tell their own investors - but even the old $50 billion rule of thumb prices the expansion near $350 billion, roughly two and a half times Meta's entire full-year capex guide.
So how does a company guiding to $145 billion double its capacity when the street math says the addition alone costs $350–700 billion? There are three ways to square that circle, and each tells a different story.
First: Meta builds for less - partly for real, partly on paper. Divide this year's capex by this year's deployments and the implied cost lands near $20 billion per gigawatt. Some of that discount is genuine engineering: Iris, the AMD gigawatts, self-built data centers and locked-in components all cut the cost of owned compute, at Nvidia's expense. But some of it is accounting. As Morgan Stanley detailed in June, headline capex understates the real commitment: purchase obligations, leases that haven't yet commenced, and rented third-party compute all keep costs off the books until delivery. Part of Meta's apparent bargain is simply the bill sitting on someone else's balance sheet - or parked in construction-in-progress, waiting to land as depreciation. And the rented slice doesn't dent Nvidia at all: the multi-billion-dollar CoreWeave and Nebius deals Meta signed are Nvidia GPUs on somebody else's books.
Second: the guide just keeps ratcheting. It has already moved from $115-135 billion in January to $125-145 billion in April, when CFO Susan Li blamed "higher component pricing" - and doubling capacity into a memory shortage is a standing invitation for hike number three.
Third: "fourteen gigawatts" turns out to be an elastic unit - contracted versus energized versus deployed - and the memo never says which.
The first path erodes Nvidia's claim on every AI dollar while confirming the off-balance-sheet worry; the second erodes Meta's free-cash-flow story; the third merely defers the question to the earnings call. There is no version in which the memo is unambiguously bullish for the whole complex at once - which is how a chip milestone nets out to a red open.
The memo also comes after the market spent mid-June mapping how any of this gets paid for. Two weeks ago we noted Goldman's argument that 2027 hyperscaler capex estimates are "too conservative" - Goldman's base case is roughly $1.1 trillion, its upside case $1.4 trillion - alongside Morgan Stanley's tally of the financing underneath: some $570 billion of AI-related debt issuance expected this year, hyperscaler gross leverage doubling from 0.9x to 1.8x in two quarters (past the entire energy sector), and Meta credit now trading wider than the investment-grade CDX index. Morgan Stanley already models Meta's 2026 free cash flow as flat to negative. Beneath the disclosed capex sits roughly $1.8 trillion of off-balance-sheet purchase and lease commitments across the complex, with Meta among the less forthcoming - it has declined to quantify the cloud-capacity portion of its $238 billion in commitments - while stretched payables and swelling construction-in-progress balances defer a depreciation load Morgan Stanley sees taking Meta from about 9% to about 19% of revenue by fiscal 2028. That leverage is migrating into the supplier and private-credit layer - vividly illustrated by the $35 billion chip-backed Anthropic SPV Apollo and Blackstone raised in June - precisely where disclosure is thinnest. A fourteen-gigawatt target stacks on top of every one of those trends.
At the high end of its capex forecast ($135BN), META free cash flow in 2026 will be $0 pic.twitter.com/xgOeHnGEZS
— zerohedge (@zerohedge) January 28, 2026Tyler Durden Thu, 07/09/2026 - 10:35