On July 1, 2026 — the first trading day of the second half — Wall Street staged a textbook split-screen.

Meta surged roughly 10%, adding ~$98 billion in market cap in a single session, one of its biggest single-day gains of the year. Meanwhile, the semiconductor sector — coming off its strongest first half on record — fell apart in unison: Micron dropped ~9–10%, SanDisk ~11%, Intel ~8%, AMD ~6%, TSMC ~6%, and even Nvidia and Broadcom slipped ~1–2%. GPU cloud plays got it worse: CoreWeave and Nebius fell as much as ~15% intraday before closing down roughly 12% each.

The catalyst: a single Bloomberg report that Meta is building a cloud computing business to sell its "excess AI compute" to outside customers.

Note that word: excess. For three years, the entire AI hardware bull market has rested on one article of faith — compute is always scarce. Now one of the Big Four buyers is publicly admitting it has more than it needs. The market immediately asked the most dangerous question it knows: Is the capex supercycle topping out?

We break it down layer by layer. Bottom line up front: this is not the inflection point — but it is a dress rehearsal for one. The real test arrives in three weeks.

Start with the facts. According to Bloomberg, Meta is developing an internal project called "Meta Compute," with two business models under discussion: one lets developers pay to access AI models hosted in Meta's data centers — similar to Amazon Bedrock; the other rents raw compute capacity directly, the model used by GPU cloud companies like CoreWeave. If it launches, Meta would compete head-on with AWS, Azure, and Google Cloud. No pricing, launch timeline, or customer list has been disclosed.

For Meta shareholders, the report reframed a nagging question. Through Q1 2026, the company had committed roughly $182.9 billion to AI infrastructure, with its 2026 capex guidance raised in April to $125–145 billion (from $115–135 billion), yet earnings calls had never broken out AI-specific revenue. Wall Street had been skeptical for months — Meta beat Q1 estimates by a wide margin in April and still sold off 7% that day; hyperscaler forward multiples briefly touched their lowest levels since the ChatGPT launch. An unconfirmed report repositioning that massive spend from "cost sinkhole" to "potential revenue stream" was enough to send the stock up 10% in a day. It bears noting: Meta declined to comment on the report, and Reuters said it could not independently verify the details.

But the speed of the rotation out of semis was equally striking. The chart below shows July 1 performance for the major names involved:

Worth noting: Apple, Microsoft, Alphabet, and Tesla all closed higher or near flat that day. What sold off wasn't "AI" — it was "AI hardware." Money didn't leave the market; it just changed seats.

Semiconductor bulls can summarize their entire thesis in one sentence: demand always exceeds supply. As long as that holds, Nvidia's GPUs, Broadcom's custom ASICs, and Micron's HBM can be priced at "sell everything you can make," and valuations can be set at "capacity equals revenue."

The Meta report cracked that premise for the first time. The logic chain runs like this: if Meta has so much compute it needs to sell the surplus, does that mean it overbought? If it overbought, will it buy less going forward? If even a buyer spending over $100 billion a year starts pulling back, does order growth for chipmakers peak?

That is the full architecture of the "capex inflection point" narrative. For a sector up 82% in the first half — the VanEck Semiconductor ETF's strongest H1 since the fund's inception in May 2000 — any signal pointing toward a top is enough to trigger an avalanche of profit-taking. Micron is up more than 260% YTD; SanDisk more than 750%. In Q2 alone, Micron, Intel, and AMD combined added roughly $2 trillion in market cap. With gains like that, nobody wants to be last out the door.

The GPU cloud companies took the hardest hit, and it's not hard to see why. CoreWeave's entire business model is buying Nvidia cards and renting them to compute-starved customers — and Meta is one of its largest clients: a $21 billion long-term contract running through end of 2032, on top of a $14.2 billion agreement signed last September. Nebius holds a separate contract with Meta worth up to roughly $27 billion over five years. When your biggest client announces it's entering your business — deploying "excess" capacity and a near-unlimited balance sheet to undercut your pricing — a 12% drop could almost be called restrained.

In the rush to panic, the primary source tends to get lost. Here is what Zuckerberg said at the shareholder meeting in May: companies come to Meta almost every week asking either for API access or to buy compute outright — and they are willing to pay above Meta's own procurement cost.

Sit with that sentence for a moment. If the industry's compute were genuinely in surplus, would buyers be lining up and bidding above cost? No. Real gluts don't look like this. Fiber optic in 2000 and housing in 2008 showed us what true oversupply looks like: assets sitting on books, marked down, still can't move. Today, GPU compute carries a positive premium in the secondary market. That is evidence of scarcity, not surplus.

Then there is the second-order implication. Zuckerberg said explicitly that being able to monetize surplus capacity "partly gives us the confidence to invest in building it." Translation: when compute shifts from pure cost to monetizable asset, Meta's incentive is to build even more, not less — the exact opposite of an inflection point. If hyperscaler capex used to be a war of attrition, compute-as-a-service turns it into a capital-intensive business with a built-in return mechanism — like an airline buying more planes because it can charter them in the off-season.

Third layer: this is already an industry trend, not an isolated case. According to CNBC, Elon Musk's SpaceX has already been selling excess compute this year: ~$1.25 billion per month to Anthropic, and ~$920 million per month to Google. Note the price signal again: if compute were truly in surplus, would buyers be willing to pay billion-dollar monthly tabs? In other words, the market is redrawing the profit map of the AI supply chain — shifting from the "shovel sellers" (chip companies) toward "compute landlords" (those who own the power, land, and data centers). July 1's bifurcated session was essentially the market's first attempt at pricing that redistribution.

There is a less glamorous but possibly more accurate read on July 1: it was Q3's first trading day — the single most natural moment for institutional rebalancing and locking in H1 gains. And the Meta report was not the only thing hitting semis: Bank of America simultaneously warned that AI trades are accumulating "bubble risk," and a meaningful portion of Intel's, AMD's, and TSMC's declines can be attributed to that note.

A few data points for context. The Nasdaq fell just 0.66% on the day; the S&P dropped 0.22%. Blue-chip names in the Dow attracted steady inflows, and the Russell 2000 small-cap index hit an all-time high, extending a six-session winning streak. Some traders called this the "Great Rotation" — out of crowded mega-cap tech, into the broader market. That is the signature of a maturing bull market, not a meltdown.

So our attribution for the July 1 selloff: seventy percent profit-taking and quarterly rotation, thirty percent narrative fracture. A 10% drop in Micron sounds alarming, but against a 260% YTD gain it simply erased part of June's final leg higher. The real information content is not the size of the drop — it is that markets just ran a live drill: when the words "compute glut" appear, this is how money moves and where it goes.

Capex inflection points do not arrive via Bloomberg reports. They arrive in numbers. Three things to watch:

Signal one: Q2 guidance from the Big Four. Microsoft, Alphabet, Amazon, and Meta together represent roughly $700 billion in 2026 capex commitments — a figure revised upward multiple times during the year. Barclays U.S. equity strategy head Venu Krishna has argued the capex peak is "not visible" within the next two to three years, and that Wall Street's 2027–2028 estimates are "clearly too low." Whether he is right gets its first real test this earnings season: Intel reports July 23 after the bell, with the four hyperscalers following in late July and early August. If any one of them cuts 2026 or 2027 guidance, the inflection narrative upgrades from "dress rehearsal" to "live event."

Signal two: GPU rental pricing. Zuckerberg's "above procurement cost" comment is the most important anchor right now. The moment spot pricing in the GPU cloud market falls below the cost of ownership, supply has genuinely caught up with demand, and "glut" moves from talking point to economic reality. CoreWeave's stock price and booking data are real-time proxies for this indicator.

Signal three: the depreciation accounting debate. The core bear argument is that this infrastructure buildout rests on rapidly depreciating chips, and that trillion-dollar investments may not be vindicated by end-user revenue. The bears get their evidence when companies start shortening GPU depreciation schedules or booking material asset impairments. That evidence has not yet appeared.

Back to the title question: has the capex inflection point arrived? Our answer: no — but the market just ran a full dress rehearsal of what happens when it does.

On semiconductors, we remain cautiously constructive: the demand-side primary evidence — buyers paying above cost for compute, the Big Four's ~$700 billion in committed spending, monetizable compute actually reinforcing investment appetite — still outweighs the supply-glut hypothesis, and July 1's decline looks more like healthy decompression after a historic first half than the start of something worse. But "cautious" carries real weight here: the sector is up 82% in six months, and every position holder has to answer one question: the multi-year industry thesis and buying at today's prices are two different things. On the former, we are not worried. On the latter, the margin of safety is narrowing.

At its core, this story is not about whether AI infrastructure gets built. It is about who collects the profits from building it. As Meta shifts from compute buyer to compute seller, pricing power along the supply chain is drifting from pick-and-shovel chip plays toward compute landlords. Long-term chip demand has not been disproved — but the era of chipmakers monopolizing outsized margins may be entering its second half.

Our invalidation conditions: if the Q2 earnings season shows all four hyperscalers raising capex guidance again, and GPU rental rates hold a positive premium, then July 1 goes in the textbook as a classic misread and the dip is a buying window. Conversely, if any one of them cuts guidance — or rental rates fall through the cost-of-carry floor — we retire the "dress rehearsal" framing and move to defense. The bell has not rung. But the warning shot has been fired. The verdict arrives in three weeks.