Three flags, one substrate

Three weeks ago the sovereignty debate was a slide. As of last Wednesday it is a rack diagram, an export approval, a merger announcement, and a postponed deadline. The artifacts arrived in clusters rather than in sequence, which is what made the past twenty-one days feel different from the previous twelve months of talk.

I want to walk through four of those clusters (Brussels, Hangzhou, Bengaluru, and what I'll call the middle powers) and then come back to a question nobody seems willing to ask out loud.

1. Brussels: delaying the rules, accelerating the racks

On 7 May the Council and Parliament closed an omnibus deal that pushes most high-risk AI Act obligations out to 2 December 2027 for standalone systems and 2 August 2028 for product-embedded ones. The Commission framed it as simplification. Read it as triage: the AI Office is under-resourced, the conformity-assessment pipeline does not exist at the scale the original calendar implied, and the high-risk-classification guidelines published on 19 May showed exactly how many edge cases were still open. A delay was the only honest move. The same omnibus also slipped in a new Article 5 prohibition on AI systems used to generate non-consensual intimate imagery, effective 2 December 2026. That is what political deal-making produces: some timeline relief, some fresh substantive obligations.

Twenty days later the same Commission is expected to table the Cloud and AI Development Act, the centerpiece of its Tech Sovereignty Package. CAIDA has two stated goals: tripling European data-center capacity by 2030, and creating EU-native infrastructure obligations for AI workloads above defined risk and scale thresholds. Those sit alongside CNBC's reporting that Brussels is weighing restrictions on US cloud platforms for sensitive government data. The picture is clearer than the rhetoric: the Commission is buying time on risk-classification and pushing harder on infrastructure. Those are two different fights. The first needs auditors. The second needs land, power, water, and silicon Europe currently imports.

The substrate question gets quieter the closer you get to it. CAIDA can mandate that an EU bank's risk model run on EU-controlled hardware. It cannot wish the GB300 racks into being. The two commercial proof points this week, SAP's EU AI Cloud and TCS's SovereignSecure Cloud, both launched on 26 May, are useful, but neither is a chip vendor. Both are operators sitting one layer above the part of the stack Europe doesn't make.

2. Hangzhou: a frontier model with no Nvidia in the loop

DeepSeek released V4 in late April at $3.48 per million output tokens for the V4-Pro tier. Set the price aside. The structural news is that the model was engineered to train and serve on Huawei's Ascend supernode, which Huawei calls "day zero" support. A frontier-grade Chinese model that doesn't require Nvidia silicon to inference is a different artifact than another open-weights release.

Be careful here. Huawei's headline figure of $12 billion in expected AI-chip revenue this year, up 60% year-on-year is a vendor claim and should be treated that way. Revenue projections live or die on yield ramp, HBM availability, and the willingness of Chinese hyperscalers to absorb supernode integration costs. The Trump administration has in parallel cleared roughly ten Chinese firms, Alibaba, Tencent and ByteDance among them, to buy up to 75,000 H200 units each. So China now has a partial Nvidia channel reopened and a credible domestic path for the first time. Beijing's communiqués call that a sovereignty win. Optionality is the better word for it, and optionality is what matters to capex committees and risk officers.

The mirror move on the U.S. side is the AI export-promotion track inside the Trump AI Action Plan: full-stack American AI packages for allied buyers, location-verification features on the chips themselves, harder end-use monitoring. Read together with the H200 carve-outs, that describes a price-tiered access regime rather than a coherent strategy of containment. Allies get the new stack with assurance. Adversaries get last-gen silicon under quotas. The middle gets to choose, but the choosing has a cost. That cost is what every other capital in this piece is trying to price.

3. Bengaluru: the launch event that was also the entry event

At the India AI Impact Summit, Sarvam unveiled 30B and 105B parameter models, BharatGen shipped Param 2, and Gnani.ai showed vertical models. The framing was sovereign stack. In the same week, OpenAI announced OpenAI for India with named local partners. India's policy posture has long preferred coexistence over exclusion, and this is what coexistence looks like when both sides have something the other wants. The Indian government wants a credentialed model layer it can shape. OpenAI wants distribution into a market it cannot afford to leave to domestic incumbents and to Chinese open-weight imports.

What is not yet visible is whether the Sarvam/BharatGen stack will run into the chicken-and-egg problem Mistral has lived with: world-class teams, real models, and a customer base that still routes high-stakes workloads to whichever foundation model offers the lowest per-token cost at the required quality. India's $200 billion AI investment figure is a forecast rather than a commitment. I would want to see Sarvam's enterprise pipeline, rather than its government anchor tenancy, before pricing in sovereign-stack dominance.

4. The middle powers

The most underdiscussed deal of the past month is Cohere's acquisition of Aleph Alpha, closed late April at a $20 billion post-merger valuation, with a $600 million Schwarz Group commitment seeding the next round. A Canadian model lab buying a German one is what middle-power consolidation actually looks like. The same pattern is visible in Saudi Arabia, where HUMAIN's $100 billion+ commitment (11 data centers, 2,200 MW of planned capacity, several hundred thousand Nvidia GPUs) converts oil revenue into compute, and in the UAE's $5 billion Digital Sovereignty Fund underwriting AI parks and regional fab. Add the European Investment Fund's €15 billion fund-of-funds, targeting roughly 100 European VC funds, and a pattern emerges.

These are not three sovereign stacks. They are three different bets on what sovereignty actually buys. Saudi Arabia and the UAE are buying physical capacity. Cohere and the merged Aleph Alpha entity are buying a regulated-customer wedge between hyperscaler frontier labs and bespoke government deployments. India and Europe are buying time and policy leverage. Only China is buying out of the dependency entirely, which is uncomfortable for everyone else, and it is doing so only because export controls forced the question early enough to matter.

The supplier line in the bill of materials

The bill of materials across all four clusters comes down to one supplier line. Nvidia shows up in three of them. In the fourth, the substitute is Huawei Ascend running a Chinese-trained model on Chinese-fabbed silicon. That is sovereign for Beijing and for Riyadh-adjacent customers, and it is also a single-vendor concentration that will become its own story the first time a node generation fails or a tape-out slips.

I doubt the cleanest framings on either side survive contact. The European version, we will mandate sovereign clouds and they will materialize, depends on a chip supply chain Brussels does not own. The Chinese version, Ascend plus DeepSeek breaks U.S. compute leverage, depends on Huawei holding HBM yield and shipping at the volume DeepSeek's pricing implies. The Indian version is the most honest of the three because it admits the optionality openly: build a domestic credentialed layer and import frontier capability from whoever sells it cheapest at the required quality.

Over the next ninety days the CAIDA draft is the document to read, and specifically whether it contains real procurement teeth or only disclosure obligations. The legal text will matter more than the press conference. Then there is DeepSeek V4 throughput on Ascend supernodes under sustained enterprise inference, where benchmarks are press releases and MFU under real traffic is the test. And there is Sarvam, which needs a meaningful enterprise contract that does not depend on Indian-government anchor tenancy.

The talent line will quietly decide more of this than the chip line will. Sarvam's 105B model exists because a specific cluster of researchers chose to stay in or return to Bengaluru. Aleph Alpha's value to Cohere is the customer-facing engineering bench in Heidelberg more than the IP. HUMAIN's $100 billion will buy GPUs faster than it buys the people who know how to operate them at MFU. Capital is the easy variable across all four clusters. People remain the constraint, and people respond to currency, weather, schools, and visa policy. A sovereignty package can legislate none of that.

Sovereignty will not be settled by a regulation or a chip. It will be settled by which jurisdictions can compound capability across three or four product cycles without losing their talent or their balance sheets. That list is shorter than the past four weeks of announcements suggest.

Tarry Singh is the founder and CEO of Real AI, an enterprise AI advisory and deployment firm working with global enterprises on production agent systems, model risk, and AI sovereignty strategy. He also leads Earthscan, an Energy AI startup, and is a founding contributor to the EU-funded HCAIM and PANORAIMA programmes for responsible AI education across European universities. He writes at tarrysingh.com.