ASML’s earnings say the quiet part out loud: AI still runs on machines

The interesting part is the market’s restraint. ASML’s outlook improved, but the share price reaction looked muted because the second-quarter guide of EUR 8.4 billion to EUR 9.0 billion sat below the average Bloomberg analyst expectation, and because the shares had already risen about 40% this year before the release. That is a useful lesson for investors. A strong business does not always produce a dramatic stock move, especially when the market arrives at the party early and has already eaten half the snacks. In other words, the question is no longer whether AI spending is real. It is how much of that good news is already priced in.

From one chip to a whole estate

The Meta-Broadcom news helps explain the next stage. Broadcom announced a multi-year, multi-generation partnership with Meta to support its custom Meta Training and Inference Accelerator, or MTIA, chips through 2029. The first phase exceeds 1 gigawatt and includes the rollout of what the companies call the industry’s first 2 nanometre AI compute accelerator, plus Broadcom’s Ethernet networking technology to connect Meta’s growing AI clusters.

Why does this matter beyond Broadcom? Because Meta is not building AI with one supplier or one type of chip. It is assembling a full estate of infrastructure: custom silicon, external cloud capacity, networking, and multiple hardware partners. Reuters has reported Meta’s expanded cloud and infrastructure arrangements with CoreWeave, Oracle, Google and AMD, alongside its own chip roadmap. That tells investors something important. The AI trade is broadening. The value pool is no longer just about who sells the headline processor. It is increasingly about who enables throughput, lowers cost, improves networking, and gets systems working together at scale. That is a much wider field.

The buildout is real, but so is the trade risk

This is where ASML and the wider AI infrastructure trade meet. ASML says customers are raising short and medium-term demand, and the company expects to ship 60 low numerical aperture extreme ultraviolet machines in 2026, with capacity for 80 in 2027. That supports the idea that the buildout is still running hard. But it also shows where the next tension sits: not demand, but delivery. If the best tools, packaging, memory, power and networking all need to arrive at once, the chain only moves as fast as its slowest piece.

Then there is trade. ASML says its 2026 guidance already allows for different export control outcomes, yet China remains a live variable. China made up about a third of ASML’s sales in 2025 and is expected to fall towards 20% in 2026 under existing restrictions, while new United States proposals could tighten things further. That matters because strong demand can sometimes be flattered by customers pulling orders forward before rules change. Investors need to separate genuine long-cycle demand from a temporary rush to buy before the door narrows.

The risks are not mysterious. The first is valuation risk. Great companies can still disappoint if expectations run too far ahead of delivery. The second is policy risk, especially for companies exposed to China or cross-border semiconductor equipment flows. The third is return risk. AI infrastructure spending can stay strong and still leave some investors unhappy if the industry builds capacity faster than end demand can absorb it. The market is now asking a harder question than last year: not who is spending, but who earns a durable return on that spending.

Investor playbook

• Watch bottlenecks, not just headlines. Tools, packaging, memory, networking and power now matter as much as the chip itself.
• Separate demand from pull-forward. Export controls can make one quarter look stronger than the true trend.
• Follow workload economics. Custom chips matter most when they lower cost for repeated tasks like inference.
• Focus on the chain, not one hero. This phase looks broader and less winner-takes-all than the first AI rush.

Behind the curtain of the AI boom

The neat story in AI used to live on the screen. A model answered a question, wrote a paragraph, or made a picture, and investors could pretend the magic happened there. ASML’s earnings and Meta’s Broadcom deal say otherwise. The real action sits behind the curtain, in lithography tools, networking fabrics, custom accelerators, data centres and export licences.

That is why these updates matter. They show the AI trade is still alive, still expensive, and increasingly industrial. For long-term investors, the lesson is simple. Do not just ask who has the smartest model. Ask who owns the bottleneck, who earns on the buildout, and who still looks sensible once the excitement leaves the room.