The Death of the GPU Shortage: How Global Funds Are Unwinding Hottest AI Trades

When everyone leaves a crowded party at once, there’s a certain kind of silence. The table still has the drinks on it. The music continues to play. But the energy has drained out of the room so fast it leaves you wondering whether the party was ever as good as it felt. That is roughly what happened to Asia’s semiconductor markets this past week, and watching it unfold from the outside, it’s hard not to feel a strange mix of vindication and unease.

The reasoning seemed sound for months. The world needed more artificial intelligence. More chips were required for AI. Chips came from Samsung, SK Hynix, Taiwan Semiconductor. Therefore, buy Asia. Global funds did exactly that, pouring money into Korean and Taiwanese equities with a conviction that felt almost religious. The KOSPI had become the world’s best-performing index.

TSMC was a darling. The GPU shortage — the very bottleneck that had throttled AI ambitions across Silicon Valley and beyond — seemed like a permanent tailwind for anyone supplying the hardware. Then, in the span of roughly two weeks, overseas investors sold approximately $16.8 billion worth of Korean and Taiwanese shares combined. The celebration came to an end.

Samsung had its worst two-day performance in nearly fifty years, falling nearly twenty percent in just two days. SK Hynix came next. In a single week, TSMC lost more than 5%. On Tuesday alone, the Korean won fell 3.3% versus the US dollar, marking the largest closing-session loss since the gloomy days of 2009. These are not typical corrections. When big, astute investors quit arguing and start fleeing, these kinds of actions take place.

Depending on who you ask, the geopolitical shock emanating from the Middle East was the immediate trigger. Growing concerns about Iran, the potential for a wider conflict, and the threat of a spike in oil prices all hit markets that were already overburdened by excitement. According to Sydney-based portfolio manager Matthew Haupt of Wilson Asset Management, crowded long positions in AI and related assets “were sold aggressively in the race to bring down exposures.” In other words, the most convenient product to sell was the AI trade.

It’s possible that the selloff is precisely what it appears to be: a brief but violent repricing, the kind of correction that eliminates excess and prepares the market for recovery. Kerry Craig of JPMorgan Asset Management stated as much on Bloomberg, pointing out that investors might return if things get better. That’s the hopeful interpretation. But there’s another interpretation, one that the naysayers have been quietly holding for over a year: that the AI rally had outrun its fundamentals, that enthusiasm for chips and data centers and GPU clusters had priced in a future that hasn’t arrived yet, and that the first serious shock was always going to expose that gap.

The AI infrastructure story’s numbers are truly astounding. NVIDIA reported data center revenue exceeding $26 billion in a single quarter last year, driven almost entirely by demand for GPUs used to train and run AI models. Industry projections put the global GPU-as-a-service market above $50 billion by the end of this decade.

These figures are real, and they reflect a real shift in how technology companies are spending their capital. The question — the one that linger in every analyst note and investor call — is whether that spending translates into profits quickly enough to justify the valuations the market assigned.

There’s a sense, watching this week’s events, that some of the largest funds in the world made a quiet decision: they didn’t need to answer that question right now. Better to reduce exposure first, reassess later, and let someone else hold the risk through the volatility. That is a rational calculation, even if it produces irrational-looking market moves. And when everyone makes that calculation at the same time, you get a week that looks like this one.

What makes this moment interesting is where it lands for companies further down the infrastructure chain. Applied Digital, trading on the NASDAQ as APLD, has been building HPC data centers specifically for AI workloads — a 400 megawatt campus in Ellendale, North Dakota, backed by NVIDIA equity and nearly a billion dollars in capital from Macquarie.

The company recently pivoted away from its GPU cloud services segment, a move that rattled short-term investors but arguably cleaned up the story for anyone willing to look past the next quarter. Whether that thesis survives a sustained pullback in AI sentiment is still unclear. It’s a company in the right industry at a genuinely uncertain moment.

The broader market for GPU compute — the infrastructure that powers everything from large language model training to real-time inference — has spawned an entire ecosystem of marketplaces and platforms trying to bridge the gap between supply and demand. The GPU shortage that defined 2023 and early 2024 may be easing, but the demand curve hasn’t flattened. It’s still climbing, just more slowly than the most euphoric projections suggested. That gap — between a market cooling and a need that remains very real — is where the next chapter of this story will be written.

For now, the exits are crowded. The Korean won is weak, the chip indices are bruised, and somewhere in Sydney and Singapore and New York, portfolio managers are staring at spreadsheets and wondering if they sold early or right on time. That uncertainty, more than any single data point, tells you exactly where we are.