Compute & Infrastructure

Broadcom's two-year AI chip sales forecast, reported by Bloomberg, positions the company as the most credible near-term challenger to Nvidia in the accelerator market — primarily through custom ASIC work for hyperscalers including Google and Meta. The company's fiscal 2028 earnings projection of over $30 per share outpaces Wall Street consensus, suggesting the revenue ramp from XPU deployments is being treated internally as high-confidence rather than aspirational.

The strategic read here is that hyperscalers are increasingly willing to absorb the upfront engineering cost of custom silicon to escape Nvidia's pricing power and supply allocation leverage. Broadcom's model — designing ASICs to customer specification rather than selling merchant silicon — insulates it from GPU commoditisation risk while locking in multi-year design wins. The risk is execution: custom silicon programs are long-cycle, and a customer cancellation or redesign can materially shift revenue timing.

Two assessments published this week converge on a damning verdict for China's domestic semiconductor tooling ambitions. Zeiss Group — the German optics manufacturer that is a critical supplier to ASML and therefore a direct participant in the EUV supply chain — stated that China is approximately 15 years behind in cutting-edge chipmaking tools, according to Bloomberg. Separately, Tom's Hardware reported analyst assessments placing China's EUV capability at roughly where ASML stood in 2004, and noted there is no confirmed evidence of Chinese vendors producing immersion lithography scanners at volume.

The critical distinction infrastructure analysts must draw is between the static gap and the rate of closure. A 15-year gap measured today against a moving frontier is not the same as a 15-year fixed delay. ASML spent two decades developing EUV with billions in coordinated European industrial investment; China is attempting to compress that timeline under export-control pressure. The Zeiss assessment is notable precisely because it comes from inside the supply chain rather than from a policy advocacy position, lending it higher credibility than government estimates on either side.

Commerce Secretary Howard Lutnick, speaking at Semicon Taiwan, confirmed the Trump administration is actively weighing a new round of semiconductor import tariffs aimed at incentivising US-based manufacturing, per Bloomberg. Notably, Bloomberg's on-the-ground reporting from Semicon indicates that supply chain companies remain broadly optimistic, citing the strength of US AI chip demand as a sufficient buffer against tariff disruption.

This optimism warrants scrutiny. The supply chain for advanced semiconductors is not amenable to rapid geographic relocation — TSMC's Arizona fabs remain years behind its Taiwan nodes on yield and process maturity, and packaging capacity in the US is structurally thin. Tariffs on imported chips would raise input costs for US-based AI infrastructure builders in the near term, even if the long-term goal is onshoring. The policy is still at the 'weighing' stage — this is an announced consideration, not a confirmed action — but the directional signal is unambiguous and procurement and contracting teams at hyperscalers should be stress-testing scenarios.

DigitalBridge CEO Marc Ganzi, in a Bloomberg interview, stated unequivocally that power — not chips, not capital, not software — is the real bottleneck in AI infrastructure buildout, per Bloomberg. Ganzi's framing that data centers are 'not the villain' reflects industry pushback against increasing regulatory and environmental scrutiny, but his identification of power as the constraint is operationally significant coming from an infrastructure investor with capital deployed across the stack.

This aligns with on-the-ground evidence: grid interconnection queues in Northern Virginia, Dublin, Singapore, and other tier-one data center markets now run two to five years in some jurisdictions. The constraint is not just generation capacity but transmission infrastructure and permitting timelines. This is driving hyperscalers toward co-location with power generation assets — nuclear, gas peaker, and dedicated renewable offtake — as the only reliable path to gigawatt-scale deployments within relevant planning horizons.

Enflame Technology's IPO drew 4,073 times retail oversubscription, completing the public listings of all four major Chinese AI chipmakers — the so-called 'four little dragons' — per Bloomberg. Tencent's backing of Enflame is strategically significant: it mirrors the hyperscaler-to-chip-vendor relationship visible in the US, where cloud giants either invest in or acquire custom silicon capability to reduce dependency on external suppliers.

The extraordinary retail demand reflects both domestic investor enthusiasm for AI hardware and the scarcity of investable semiconductor equity in China following delistings and restrictions on foreign capital access. From an infrastructure standpoint, the completion of these IPOs means the Chinese domestic compute ecosystem now has publicly capitalized hardware vendors at scale — providing balance sheet capacity for R&D and manufacturing investment independent of state funding alone.