The Shortage Is Spreading Sideways: CPUs, Tools, Even People
AMD says server CPU demand was unforecasted, ASM sees HBM driving memory higher, and Google's AI reshuffle shows talent follows compute
The GPU shortage was chapter one. Now it is CPUs, fab tools, and people.
For two years the AI bottleneck was GPUs. On August 4 and 5, three separate signals showed it has spread sideways. AMD said server CPU demand was unforecasted and the supply chain is tight. ASM International said HBM is pulling memory sales higher in the second half. And Google reshuffled its top AI leadership, a reminder that when compute is scarce, talent moves to where the compute is.
This matters because unforecasted CPU buying means hyperscalers are adding general compute above plan to feed inference, and record tool orders are how a shortage turns into fab spending. That spending is 2028 supply, and the classic signal to watch for a cycle top in two years.
1. CPUs: demand hyperscalers did not forecast
On AMD's Q2 2026 call on August 4, CEO Lisa Su was asked if server CPU supply was constrained and whether it lasts into 2027.
Her answer, per Benzinga's transcript: "The server CPU supply chain is tight right now, it has been tight for the first half of the year because much of this demand was unforecasted."
She added that 2027 should be better because demand is now better forecasted, and that AMD feels "very good" about more than 70% year over year growth in its server CPU business, with room to go higher depending on how things play out.
Why this is a tell: CPUs are not the AI accelerator, they are the host. When inference scales, every GPU needs CPUs, memory, and networking around it. If hyperscalers under-forecast CPUs, it means inference workloads grew faster than their own capacity plans. That is demand pulling general compute, not just accelerators.
It also helps the laggard. Counterpoint noted in May that Intel's own supply constraints helped AMD gain x86 server share. When the whole CPU chain is tight, scarcity lifts pricing and share for whoever can ship.
2. Tools: HBM orders are turning into fab spend
A day earlier, ASM International reported Q2 2026 results. Revenue crossed EUR 1 billion for the first time, up 24% year over year at constant currency, with record ALD sales and spares up 34%.
On memory, CFO Paul Verhagen said on the call: "Memory sales increased sequentially compared to Q1 and were mainly driven by HBM-related DRAM applications."
And on the outlook: "We expect memory sales in the second half to be substantially higher than in the first half, driven by strong demand for advanced DRAM solutions."
CEO Hichem M'Saad added that 2 nanometer remains the largest driver, 3 to 7 nanometer is seeing an uptick again on demand for advanced CPUs and agentic AI workloads, and 1.4 nanometer pilot lines should start contributing in the second half of 2026.
The translation: toolmakers are not just talking about HBM tightness, they are booking it. Memory was 15% of equipment sales in the first half, slightly below 2025, but guided to be substantially higher in the second half. When deposition tool orders rise on HBM DRAM, that is fab capex being committed today for supply that arrives in 2028. It is bullish for semicap now and the supply to watch later.
For context on why HBM matters so much to the memory cycle, see Memory Is Not a Commodity: Why MU and SK Hynix Are Priced for a Recession That Isn't Coming and NVDA Rubin Economics: Why $8.3M per Rack Still Prints 78% Margin at $53/GB HBM.
3. People: talent follows compute
On August 5, Google announced a major AI leadership change on its blog. Demis Hassabis moves from CEO of Google DeepMind to Chair of Google DeepMind and Chief Scientist of Alphabet to focus on AGI and science, while CTO Koray Kavukcuoglu becomes SVP of Google DeepMind. Chief Scientist Jeff Dean is leaving after 27 years with Sanjay Ghemawat to launch an independent public benefit corporation, Discovery Loop, focused on ML and science, per CNBC.
As @benitoz put it yesterday in a post with 127K views on the shakeup: "Compute scarcity and corporate politics are a terrible combination. In the AI race, talent follows compute. Always."
The point is structural. When compute is the binding constraint, researchers go where they can get it. Google said it will be a founding investor and Cloud partner to Dean's new venture, which keeps the compute link intact even as talent decentralizes. For more on how power and compute have become the gating factors, see Power Is the New GPUs.
The deep insight: unforecasted means inference is pulling everything
The word to focus on is "unforecasted."
Hyperscalers plan CPU capacity a year ahead. If first half 2026 CPU demand exceeded those plans, it means inference growth surprised to the upside. Inference is not just GPUs, it is full servers. That pulls CPUs, DDR5, HBM, networking, and power at the same time.
ASM's guide confirms the second leg. When memory sales inflect on HBM in the second half, tool orders follow. Those tools build the 2028 wafer supply. Every memory supercycle ends the same way, with too much supply ordered at the peak. The time to watch for that top signal is not now, it is when those orders start to get cut, likely in 2027 to 2028.
Impacted: who benefits and what to watch
| Ticker | Price Aug 6 close | Why this matters |
|---|---|---|
| AMD | $492.18 | Claimed over 70% YoY server growth, tight supply supports pricing and share gains |
| INTC | $100.20 | Scarcity lifts the laggard, distributors reportedly filling only ~40% of allocations earlier this year |
| ASML | $1,712.74 | EUV and ALD intensity rises at 2nm and 1.4nm, benefits from logic and HBM DRAM capex |
| AMAT | $529.04 | Deposition and etch leverage to HBM and advanced logic capacity adds |
| LRCX | $307.78 | Etch and memory exposure, benefits as DRAM transitions to 4F squared and HBM |
Prices as of Aug 6, 2026 close via Bargo SIP feed.
What to watch next
- AMD supply in Q3: Does tightness ease or does 70%+ growth get capacity constrained. Q3 guide is ~$13B revenue, plus or minus $300M.
- ASM memory mix in Q3: Does second half memory actually exceed first half by the guided substantial margin, and do 1.4nm pilot orders materialize.
- Hyperscaler capex: If CPUs were unforecasted, watch for upward revisions to server capex, not just GPU capex.
- The 2028 tell: Tool order books are bullish now. The bear signal later will be order pushouts or cancellations when HBM supply catches up.