Sovereign Compute, Splintered Silicon: How the Global AI Hardware Stack Fractured in 2026

The Biden administration's AI Diffusion Rule was supposed to be the framework governing where advanced AI chips could legally go for years to come. It lasted four months from publication to rescission, and was killed two days before its own compliance deadline. What replaced it wasn't a simpler rule — it was a series of bilateral deals, summit-level trades, and country-specific waivers, layered on top of a foundry market that consolidated further around Taiwan rather than diversifying away from it. Here's the fracture, in two parts: the policy collapse, and the geography underneath it that didn't move the way policy intended.

From blanket rule to bilateral deal-making

The Biden administration issued the AI Diffusion Rule on January 15, 2025 — a tiered, worldwide licensing regime for AI chip exports, with compliance required by May 15, 2025. It never took effect. The Trump administration's Bureau of Industry and Security (BIS) rescinded it on May 13, 2025, two days before the compliance deadline would have kicked in. The Oct 2023 and Dec 2024 advanced-computing export rules — narrower, chip-specific controls — remained in effect; it was specifically the tiered, country-bucket framework that got scrapped.

What replaced it wasn't nothing. In May 2025, BIS issued guidance stating that using Huawei's Ascend 910B, 910C, or 910D chips "risks violating US export controls" — a targeted warning rather than a blanket rule. Then came the deal-making: after an October 30, 2025 meeting between Trump and Xi in Busan, a one-year waiver on a "50% rule" took effect November 10, 2025, traded in exchange for a delay in Chinese rare-earth export restrictions. On December 8, 2025, a separate one-year waiver allowed H200 chip exports to China. Read together, these aren't exceptions to a rule — they're the actual policy mechanism now: access to advanced AI chips negotiated bilaterally, summit by summit, against unrelated leverage like rare earths, rather than governed by one publicly legible framework anyone could plan five years around.

TSMC's grip tightened, it didn't loosen

If the policy goal, implicit or explicit, was to diversify chip manufacturing away from a single geopolitically exposed point, the 2025 data runs the other direction. TSMC's share of the global foundry market rose to roughly 69.9% of full-year 2025 revenue (about $122.54 billion), with quarterly figures running even higher — around 71-72% in Q3 2025 and about 70.4% in Q4. That's up from 62.8% in Q1 2024. Samsung held roughly 7.2% (the #2 spot), and SMIC rose to about 5.3% (#3). Every point of concentration increase happened during the exact window US policy was ostensibly pushing toward de-risking — worth sitting with plainly rather than letting the "re-shoring" narrative imply the opposite happened.

Re-shoring is real, but partial

That doesn't mean nothing moved onshore. TSMC's second Arizona fab completed construction in early 2026, and a third fab — targeting 2nm/A16 process nodes — was pulled forward on the roadmap, though it's now expected around 2027 rather than sooner; I'm labeling that 2027 timeline as reported rather than confirmed, since it comes from trade-press tracking rather than a TSMC disclosure I could verify directly. The $6.6 billion direct CHIPS Act grant to TSMC is well-established; broader multi-fab investment totals circulating in the ~$165 billion range should be read as reported estimates, not confirmed figures.

Intel's 18A process reached high-volume production at Fab 52 in Arizona, with Panther Lake as its first shipping product (launched January 2026) — but as of this writing, Intel has no major external foundry customers on 18A. The company received $8.9 billion from the US government and a separate $5 billion investment from Nvidia; it's worth being precise that Nvidia's investment did not come with a commitment to actually manufacture chips at Intel. Reported yield estimates put Intel 18A somewhere around 60-75%, against TSMC's reported >90% — both figures are secondary estimates, not confirmed by either company, and I'm flagging them that way rather than presenting them as settled. On the Chinese side, SMIC continues 7nm production for Huawei with yields described as subpar but improving, while 5nm yields remain poor — again, general characterizations from trade press rather than precise verified figures.

TSMC's 2nm (N2) node reportedly entered volume production in Q4 2025 at somewhere around 90,000 to 100,000 wafers per month — that specific figure is a secondary-source estimate and should be read as reported, not confirmed.

The rest of the world is hedging too

Japan, India, and South Korea are each building out their own fab or advanced-packaging capacity in parallel to the US push — Rapidus's 2nm ambitions in Japan, Tata-PSMC's Dholera project and Micron's Sanand facility in India, Samsung's Taylor, Texas fab, and SK hynix's HBM4 development all point the same direction, though I'm deliberately not citing specific milestone dates or dollar figures for any of them here, since I couldn't verify those to the standard the rest of this piece holds. India's IndiaAI Mission GPU tender is reported at roughly $1.25 billion (₹10,371 crore) — again, reported rather than confirmed against a primary government filing.

The Middle East took a different route: buying assurance rather than building fabs. Microsoft's $1.5 billion stake in the UAE's G42 grew into a $15.2 billion total commitment through 2029. That's the corporate-silicon side of this same story, covered in more depth in the companion piece on custom AI chips and sovereign cloud deals — that post maps who's building the chips and buying the assurance; this one maps where the chips are actually fabricated and under what export rules they can move. Worth reading together, but they're answering different questions. The underlying capacity crunch connects directly to how much power this entire build-out actually consumes, a constraint that doesn't respect export-control boundaries the way policy does.

The view from Kathmandu

None of this — the Diffusion Rule, its rescission, the Busan trade, TSMC's rising share — was decided with a country like Nepal anywhere near the table, and that's precisely the point worth naming plainly. The fracture map isn't really about Taiwan versus Arizona versus Shenzhen for most of the world; it's about which countries get a tier assignment, a waiver, or an assurance deal at all, and which simply aren't part of the conversation. Silicon Valley experiences this story as a competitive-advantage question between three or four companies and two governments. From Kathmandu, the more honest framing is that the global compute map is being redrawn by negotiations most countries have no seat in — which is a bigger story than who wins the 2nm race, even if it gets a fraction of the coverage.

Sources

• Bureau of Industry and Security (bis.gov) — AI Diffusion Rule issuance, January 15, 2025; rescission, May 13, 2025
• Bloomberg — "Trump to Rewrite AI Chip Curbs," May 13, 2025
• O'Melveny; Akin Gump — BIS guidance on Huawei Ascend 910B/910C/910D chips, May 2025
• IISS — "The US pivot on regulating AI diffusion," analysis of Busan waiver (Nov 10, 2025) and H200 China waiver (Dec 8, 2025)
• Taipei Times; TrendForce — TSMC ~70% full-year 2025 foundry share; quarterly figures
• Motley Fool; TrendForce — foundry market share trajectory, Q1 2024–Q4 2025
• CNBC — Intel 18A high-volume production, Fab 52 Arizona, December 19, 2025; Panther Lake launch, January 2026
• TrendForce — TSMC Arizona third fab timeline; SMIC yield status

Written by Abhishek Kushwaha, founder and writer at Global Tech Search, based in Kathmandu, Nepal.