Article
AI News Chipmakers Broker Commentary

Chip stocks shed billions on AI glut fears that SemiAnalysis calls a misread of Meta

by TechDefused Newsroom
The image shows a silicon wafer, which has numerous microchips arranged in a grid pattern. This wafer is typically used in semiconductor manufacturing processes to create integrated circuits. — Credit: Photo by Laura Ockel / Unsplash cPhoto by Laura Ockel / Unsplash
Photo by Laura Ockel / Unsplash

A single rumour about Meta was enough to wipe billions of dollars off semiconductor stocks in a day. SemiAnalysis says the market read it exactly backwards.

The research firm called the panic over excess AI computing capacity "erroneous," in a note published after Thursday's global sell-off. Its argument is blunt. Meta is not about to dump spare chips on the world. It is buying more.

What actually spooked the market

The trigger was a Bloomberg report that Meta is building a cloud business to sell access to its surplus AI compute.

Investors took that as proof of a fear that has hung over the sector for months. If the largest buyers of graphics processing units start renting out spare capacity, the world may already have too many chips.

That one interpretation, surplus for hire, revived the AI overcapacity narrative in a single session.

How far the sell-off spread

The alternative cloud providers took the worst of it. CoreWeave fell 14% and Nebius dropped 17%.

The US chip names followed. Micron, AMD, Intel and SanDisk posted losses ranging from 4 to 14%.

The damage crossed the Pacific. Samsung Electronics slid about 9% and SK Hynix fell around 14%, while China's SMIC and Hua Hong each lost close to 10%. South Korea's Kospi triggered a trading halt as the memory makers sank.

Why SemiAnalysis says the read is wrong

SemiAnalysis rejected both popular takes, the glut story and the idea that Meta is trimming costs.

Meta's data-centre and compute procurement will accelerate, not slow, the firm argued. It expects 2027 capital spending to be, in its words, "shockingly high."

The numbers it cites are the point. In the first six months of 2026, Meta contracted more than 5GW of capacity across cloud and colocation, before counting its own self-build work. The two largest campuses under construction account for 2.5GW between them.

That figure also punctures a separate claim doing the rounds, that half of US data centre projects are delayed and only 5GW sit under construction nationwide. Two Meta sites alone are half of that.

Where every gigawatt finds a home

The core of the argument is demand, not supply.

SemiAnalysis lays out four high-value outlets for Meta's compute, each unlike the bare-metal renting the market feared.

Meta is still training frontier models, so the top tier of capacity stays in-house. It is scaling its advertising recommendation systems more than 10-fold, which needs both training and inference power and lets advertisers pay more. It is preparing a model API service in the mould of Amazon's Bedrock. And it can lease the rest on short, high-margin terms rather than commodity rates.

On the API point, SemiAnalysis reports that Meta is in late-stage talks with Anthropic to run Claude inside its own data centres as a private deployment, which Meta could then bundle for outside customers.

Every gigawatt, in other words, has somewhere profitable to go.

The panic had a split personality

The market never spoke with one voice on this.

The same cloud-resale reports that sank chipmakers pushed Meta's own shares up roughly 9% earlier in the week. Investors liked the idea of turning fixed data-centre costs into recurring revenue.

So the sell-off punished the suppliers while rewarding the buyer, on identical news.

That gap is the tell. A story cannot mean glut for the chip sector and growth for Meta at the same time, unless the market has not settled what the story actually is.

For now, the firm closest to the silicon expects more building, not less.

by TechDefused Newsroom