Meta is pivoting to cloud computing — is this the AI bubble bursting?

Bloomberg reported on 1 July that Meta is developing a cloud business, internally named Meta Compute, to sell AI computing capacity to outside customers. Meta shares rose as much as 10% intraday on the news — and the rest of the market fell in direct proportion to how exposed it was to the fear that shortage was ending.

CoreWeave closed down 13.9%; Nebius closed down 17.0% the same session, both companies carrying multi-billion-dollar, multi-year compute commitments from Meta. Micron and SanDisk fell more than 10% on Wall Street that same day. The selloff reached Asia the following session — Samsung Electronics closed down 9.1%, SK Hynix down 14.6%, dragging the broader KOSPI with them.

The fear was straightforward: if one of the hyperscalers racing to build AI infrastructure now has some left over to sell, maybe the shortage driving this entire trade is ending.

The reaction is overdone. Here is why.

Meta's own spending doesn't match the surplus story.

What would you expect a company to do if it had just discovered it built too much AI infrastructure? Slow down. Cancel a few contracts. Stop signing new ones.

That is not what Meta did.

In March 2026, Meta signed a long-term AI infrastructure agreement worth up to USD 27 billion with Nebius. The deal includes USD 12 billion in dedicated capacity and a further USD 15 billion purchase commitment, built around Nvidia's next-generation Vera Rubin platform, with deliveries beginning early 2027.

In April, Meta expanded an existing agreement with CoreWeave by a further USD 21 billion through December 2032 — on top of a prior USD 14.2 billion commitment, taking the relationship to roughly USD 35 billion. In mid-June, two weeks before the "excess compute" headline, Meta reportedly signed for 1.6 gigawatts of data centre capacity from Crusoe, across sites in Childress, Texas, and Warrenton, Missouri.

That is three suppliers in roughly four months — all committed in the weeks before Meta reportedly told the market it had capacity to spare.

There is a second detail worth considering. The Financial Times reported that Google told Meta, also around March, that it could not supply the full Gemini computing capacity Meta had requested. The shortfall was serious enough to delay some of Meta's own internal AI projects and force Meta to tell staff to use AI tokens more sparingly. Meta had been leaning on Google's Gemini for tasks like content moderation and scam detection, reportedly because it outperformed Meta's own Llama models at that work. Meta began shifting those workloads to Muse Spark, its own newly developed model, instead.

Put the two facts together. The company that reportedly told the market it has spare compute is the same company that, as recently as March, could not get enough of it from Google — hardly the picture of a company swimming in surplus.

Meta's report tells you nothing about the industry.

Even taking Meta's report entirely at face value — does what's happening at one company tell you anything about the rest of the industry?

Google's cloud backlog rose to USD 462 billion in Q1 2026, nearly double the USD 240 billion reported at the end of Q4 2025. That figure is contracted, signed revenue — not a sales pipeline or forecast. Chief executive Sundar Pichai told analysts plainly: "We are compute-constrained in the near-term... our cloud revenue would have been higher if we were able to meet the demand."

Google itself was paying SpaceX's xAI division USD 920 million a month for 110,000 Nvidia GPUs, under a contract running through 2029. A hyperscaler with its own data centres, its own chip supply deals, and one of the largest infrastructure budgets in the industry does not go out and buy capacity from another company if the industry is oversupplied. It does that because it needs more than it has.

Amazon reported the same underlying story earlier that quarter. Chief executive Andy Jassy told analysts the cost of memory "has skyrocketed" and that the industry is "just in a stage where there's just not enough capacity for the amount of demand." On chips specifically, Jassy said computing capacity built on Amazon's next-generation AI chip — still roughly 18 months from broad availability — 'has already been reserved.' Capacity on the current generation is already 'nearly fully subscribed,' only months after launch. If customers are committing to capacity that does not exist yet, the shortage extends well past this quarter's headlines.

Over in Asia, Samsung and SK Hynix pledged KRW 800 trillion — approximately USD 518 billion — to build new chip fabrication sites in South Korea's southwest. On top of previously announced projects, that brings their combined commitment to KRW 3,200 trillion, or roughly USD 2.07 trillion. Think about what that decision requires. SK Hynix's and Samsung's management have visibility into their own order books that nobody outside either company possesses.

That confidence is worth weighing against history. Both companies know the cost of getting this wrong better than most — a downturn in chip prices drove SK Hynix to the edge of bankruptcy in 2001, and both firms posted large losses as recently as 2023. They are not committing hundreds of billions of dollars because they have forgotten how badly this can go. They are doing it with that history in view.

The market treated Meta's report as evidence about the whole industry, and sold off Asian semiconductor names on that basis. The evidence says otherwise.

If demand were fading, the numbers would show it. They don't.

If there were a real slowdown, it would show up in company data — contracts renegotiated, deposits pulled back, orders cancelled. None of that is happening.

Micron holds 16 strategic customer agreements structured as take-or-pay — meaning customers must pay for the capacity whether they end up using it or not. As of its most recent quarterly disclosure, customers had already handed over USD 18 billion to USD 22 billion in cash upfront, before a single chip has shipped, to secure their place in the order book. None of that was renegotiated after Meta's report. Chief business officer Sumit Sadana told analysts the company doesn't "really see when the supply is going to be able to meet demand. That is not something we are able to project at this time."

The pricing tells the same story. On 4 June, chief executive C.C. Wei told TSMC shareholders he would "like" to raise prices, "as its competitors have done." Three weeks later, supply chain analyst Tim Culpan reported that TSMC had begun doing exactly that — rolling out increases of 5% to 10% across nearly all of its advanced manufacturing, covering nodes that make up roughly three-quarters of its wafer revenue.

Wei wasn't exaggerating about the competitors. On their own first-quarter earnings calls, Samsung disclosed that memory prices — DRAM and NAND combined — rose in the 80% to 90% range quarter-on-quarter, and SK Hynix disclosed increases in the 60% to 70% range. SK Hynix's Q1 2026 operating margin reached 72%, an all-time high. Its share price reflects it — up 235.9% year-to-date as of 2 July 2026, nearly double Samsung's 138.5% gain over the same period. Companies do not get this kind of pricing power unless customers have nowhere else to go.

Zoom out further, to the level of an entire economy. Korea's semiconductor exports rose 199.5% year-on-year in June, to USD 44.8 billion — making Korea only the fourth country in history to record USD 100 billion in total monthly exports. That is not one company's earnings call. That is a national trade statistic, recorded while the market was betting on a slowdown.

The leaders in the industry say the same thing.

Jensen Huang flew to Seoul on 7 June and told reporters the shortage would "persist for several years." Apple's own corporate statement read: "We have never seen a component price increase this much, this quickly." Chief executive Tim Cook, to the Wall Street Journal, put it more bluntly: "This is a hundred-year flood... I've never seen anything like it in over 40 years." Microsoft's own blog stated memory prices had "increased by more than 2.5 times," with "another doubling" expected by fall 2027 — and raised Xbox prices to match.

ASML chief executive Christophe Fouquet told analysts in April: "In the memory business, many customers have confirmed that they are sold out for the remaining of the year, and that they expect the supply limitation to persist beyond 2026." ASML is now committing to more than double its own 2025 output by 2027 to keep pace.

None of this looks like a slowdown.

Same demand, a fraction of the price.

The Philadelphia Semiconductor Index — the benchmark for US chip stocks — trades at 28.3 times 2026 earnings and 20.3 times 2027 earnings. The FactSet Asia Semiconductor Index trades at 14.5 times and 10.3 times over the same two years. Roughly half the multiple.

That gap isn't because Asia is growing slower. Consensus estimates put 2026 earnings growth for the Asia index at 211%, against 132% for the US index — and the two converge to almost the same rate in 2027, at roughly 40% each. The market is paying twice as much for the same growth.

TSMC — currently the only manufacturer producing the world's most advanced AI chips at scale — trades at roughly 23 times. SK Hynix and Samsung, which together control nearly 80% of global HBM supply as of 2025 and post margins ahead of Nvidia's at the company level, trade at just 5 to 6 times. Micron — smaller than SK Hynix and comparable to Samsung in HBM market share — trades at a premium to both at close to 7 times.

We continue to recommend the Global X Asia Semiconductor ETF (HKEX: 3119), which captures the full Asian semiconductor supply chain without the concentration risk of a single-market ETF. Based on a fair PE ratio of 18 times applied to 2028 earnings estimates, our target price for the ETF is HKD 334. The ETF last traded at HKD 191.7 as of 2 July 2026, which we estimate implies upside of more than 74% over that horizon.

Meta might be selling its computing power, but the evidence still points to an industry experiencing strong AI demand and a supply shortage. Asia semiconductors — the manufacturing backbone of the AI buildout — continues to trade at half the multiple of US peers, and is still worth owning.

Declaration:

This research report was prepared with the assistance of artificial intelligence (AI) tools. iFAST Financial Pte Ltd does not rely exclusively on AI for content generation; the content of this report – including all investment theses, ratings, price targets and conclusions – has been independently reviewed and verified by the research analyst(s) to ensure accuracy and professional integrity.

For specific disclosure, at the time of publication of this report, IFPL (via its connected and associated entities) holds a NIL position in the abovementioned securities. The analyst who produced this report holds positions in Global X Asia Semiconductor ETF.