The state moves into the stack

Thematic essay — week of June 28 to July 5, 2026

For two years the argument about state power over AI was an argument about regulation — rules written on the outside of an industry, applied to it from a distance. The week of June 28 to July 5 is when that framing quietly stopped describing what is actually happening. In seven days, two governments were seen doing something more direct than regulating: acquiring positions inside the AI stack. Washington gated a flagship model's release, kept the highest-capability tier behind a US-only clearance list, reinstated another model on standing conditions, and floated taking a five-percent equity stake in the company that builds it. New Delhi cleared ₹1.25 lakh crore for a second semiconductor mission, brought a third packaging plant into commercial production, signed a committed subsea cable, and reaffirmed compute subsidised to ₹65 an hour. Both are states moving from the outside of the stack to the inside. What the week makes legible is that they are moving into different layers, with opposite instruments — and that the gap between the two is the most honest available description of India's structural position. This essay is about what India's state-in-the-stack strategy actually buys, when the layers it has chosen not to own were shown, the same week, to be instruments of another state.

The two columns

Read the week as two columns of moves, and the shape resolves.

The American column is about the layers with natural chokepoints — the frontier models, and the release authority over them. On June 26, the US Commerce Department re-admitted Anthropic's highest-capability model, Mythos 5, not as a product decision but as a government-administered trust list of roughly a hundred US organisations. As the June 28 digest put it, "this is a regime with an admissions process, not an incident being unwound" — and the admission was drawn at the US border, no non-US organisation on the list. The same day, OpenAI began previewing GPT-5.6 to about twenty organisations "whose participation was coordinated with the US government." On June 30 the export-control order that had shut Fable 5 and Mythos 5 off worldwide was lifted — and Fable 5 came back on July 1 with conditions attached: pre-release cooperation on future models, ongoing risk reporting, alignment "across the US Government." The July 1 digest named the change precisely: "a US frontier lab's release calendar has a government seat at the table," and "what began on June 12 as an emergency order is ending as a structure." Then, on July 2, the Financial Times reported that OpenAI had proposed the US government take a roughly five-percent stake — about $42.6 billion at its $852 billion valuation — through an Alaska-style sovereign fund, with other frontier labs invited to cede similar stakes; the report landed while the White House finalised a voluntary pre-release review framework granting the government up to 30 days of access to a covered model before public launch, an announcement expected as early as the week of July 7.

The Indian column is about the layers below the frontier — the foundations. On June 30, the Finance Ministry's Expenditure Finance Committee cleared a ₹1.25 lakh crore outlay for India Semiconductor Mission 2.0, about 64% above ISM 1.0, sending it to the Cabinet. On July 2, a consortium of Lightstorm, Microsoft, Singtel, and Tata Communications signed contracts to build I-2SEA, a ~3,600 km subsea cable with dual east-coast landings at Machilipatnam and South Chennai, targeting service in Q4 2029. On July 1, MeitY marked eleven years of Digital India by reaffirming the IndiaAI Mission's spine — 15 supported indigenous foundation models and compute offered at ₹65 per hour. And on July 4, Prime Minister Modi inaugurated CG Semi's ₹7,500-crore packaging-and-test facility in Sanand — as the July 5 digest recorded, "the third assembly-and-test plant to enter commercial production in India this year," after Micron's and Kaynes Semicon's.

Two governments, one week, both stepping into the stack. But the July 4 digest already stated the underlying identity, in the course of covering the OpenAI equity proposal: "India's sovereign-AI programme exchanges subsidised compute for model deliverables; Washington is debating exchanging release legitimacy for equity. Different instruments, same direction — states acquiring structural positions in the AI stack rather than regulating it from outside." The essay's job is to take that sentence seriously and ask what follows from it — because the two columns are not symmetric, and the asymmetry is the whole story.

The stack, layer by layer, and who holds what

The way to see the asymmetry is to lay the stack out vertically and mark, at each layer, what India holds, the instrument it used to get there, and what it is exposed to. The layers with chokepoints sit at the top; the diffuse, capital-efficient layers sit below.

Trace the column of exposures top-down and a pattern falls out. The two layers where India is a pure taker — frontier models and leading-edge silicon — are exactly the two layers that carry chokepoints, and exactly the two that another state demonstrated it could operate as instruments this window. Everything below is where India has been building, and everything below is diffuse: a packaging plant, a data centre, a fibre pair, a subsidised GPU-hour, an application. Those layers are capital-efficient to enter and, not coincidentally, hard for anyone to switch off from outside. India's build is resilient at the layers it owns and exposed at the two it doesn't — and the two facts are the same fact, viewed from different ends of the stack.

The June 29 South Korea item drew the frontier-silicon half of this map cleanly. Korea committed fabrication-scale capital — a ~$518 billion four-fab ecosystem programme — to "own the silicon, at any cost," because Korea already sits at the fabrication frontier and is doubling down on what it leads. India's ISM 2.0, at about $13 billion, is "roughly one-fortieth" of that, and the June 30 digest named the difference in kind rather than degree: "Korea is buying the frontier it already occupies; India is buying the foundations of an ecosystem it does not yet have." The CG Semi inauguration on July 4 is what "buying the foundations" looks like in practice — OSAT packaging and test for automotive and industrial chips, the segment where Indian entry was most tractable, arriving on the timeline the mission promised. It is real. It is also, as the digest was careful to record, "several tiers above what this plant ships" from the AI-accelerator layer the podium framing invoked. Both halves are true, and holding both is the discipline.

The instrument, not the intent

The mechanism the daily format kept gesturing at, and could not stop to dissect, is why the instruments differ — and the answer is that each state reached for the layer where it already had leverage.

Washington's instruments this week all operate at the chokepoint. A release gate works because there are three or four frontier labs and they are all American; a US-only clearance list works because the highest-capability model has a single vendor who answers to a single government; an equity stake works because the labs need capital and legitimacy the state can withhold or confer. None of these instruments would function at a diffuse layer. You cannot gate the release of a data centre or clear a subsea cable entity-by-entity to a hundred approved firms. The US moved into the layers where a small number of control points make the whole layer governable — and the export-control episode was the proof of concept. As the July 2 digest put it, "model access has been demonstrated as an instrument of US policy, applied and released on Washington's timetable."

India's instruments operate where India has leverage, which is capital deployed patiently at diffuse layers: incentive outlays for packaging plants, subsidised compute-hours, consortium participation in cables, grants to foundation-model labs. These are the instruments available to a state that is not at the frontier of silicon or models and has chosen not to try to buy its way there. They are slower and their payoff is a decade out — I-2SEA lands in 2029, ISM 2.0's money does not exist until the Cabinet signs — but they compound, and they are not revocable by anyone else's directive.

The July 5 Anthropic item is the tell that the chokepoint instrument is only getting sharper. Three days after the export-control order lifted, the same company was reported to be building account-level enforcement of a geopolitical boundary — inferring jurisdiction from "time zones and usage patterns," demanding government ID and live selfies from flagged accounts. The July 5 digest read the trajectory precisely: "frontier-model access is becoming compliance-graded, verified, and revocable at the account level, not just the country level." The June shutoff was the blunt version of state leverage over the frontier layer; the telemetry enforcement is the high-resolution version. The direction is the same, and it runs one way — toward more granular control at the layer India does not own.

And then the equity proposal, which is the instrument taken to its logical end. If a release gate aligns the state with the labs episodically, a five-percent stake aligns them structurally and permanently. The July 4 digest drew the consequence for India in one line: "the counterparty behind the API now plausibly includes a shareholder state with its own geopolitical ledger." That is not a metaphor. It is a description of what an Indian bank, SI, or product startup is contracting with when it calls a frontier API — increasingly, a company whose largest and most consequential shareholder may be the government that already proved, on June 12, that it will switch the model off when it decides to.

What this looks like elsewhere

Three comparables sharpen where India sits.

South Korea — buy the frontier you already lead. Korea's move is only available to a country already at the fabrication frontier; its four-fab, half-trillion-dollar programme is a doubling-down, not a leap. India could not replicate it on the IndiaAI Mission's budget even if it wanted to, and the June 29 digest was explicit that "the comparison does not indict India's choice. It clarifies what that choice is." Korea buys the top of the stack because Korea already occupies it. India buys the foundations because it does not.

The United States — buy release authority and equity over the layer you host. Washington is not building anything this week; it is acquiring positions — clearance authority, review windows, and potentially equity — over labs it already hosts. This is the instrument of a state that sits atop the frontier layer and wants to convert incumbency into standing control. The Alaska-fund framing is the giveaway: a sovereign claim on the upside of an asset the state considers strategically its own.

India — buy the diffuse foundations you can afford. Between the two, India's instrument is the capital-efficient middle: subsidise access, fund packaging and connectivity, grant foundation-model capacity, and regulate the application layer sector by sector. It plays to a genuine strength — the model-and-application layers where the June 29 digest said India can "win on the layers where capital efficiency is possible, and to remain, at the silicon layer, a buyer."

The domestic-governance thread that also ran through the week belongs on this map as the one layer where India moved fastest, not slowest. In seven days: the Supreme Court set a zero-tolerance rule on AI-hallucinated citations and ordered the Bar Council to frame norms; the Delhi High Court sorted 52 deepfakes of an MP item-by-item, taking six down and protecting 46 as satire; the RBI's June 30 gap-assessment deadline put every regulated bank board on record with a frontier-AI risk position; and MeitY's secretary signalled that a dedicated AI statute is starting to move, saying, per PTI, "it appears that the time is getting right." Governance is the one layer where a state at India's position can act at full resolution, because it governs its own jurisdiction rather than someone else's chokepoint. It is telling that this is exactly where India was quickest — and it is worth noticing that governing the use of frontier models is not the same as controlling their supply. India is writing excellent rules for a layer it consumes.

Where this lands

The near-term signals to watch are unusually concrete this quarter, because most of the week's moves were the kind that resolve on a date.

• The White House release-standards announcement, expected as early as the week of July 7. Where the capability threshold lands decides whether the review regime catches only true frontier releases or routine model updates — and whether India's access to newest models now carries a standing US review step by design.
• Whether any non-US organisation ever enters a Mythos 5 approved set, and whether GPT-5.6 general availability includes India in the first wave. These are the cleanest single indicators of whether frontier-model access is renormalising or restratifying along national lines. The June 28 digest's formulation still stands: "frontier capability arrives in India when the US process clears it, not when the lab ships it." The GA date tests whether that hardens or relaxes.
• Cabinet ratification of ISM 2.0. The ₹1.25 lakh crore is a cleared proposal, not money; ratification converts it, a trim or stall leaves the position unchanged. The scheme guidelines — how the outlay splits between fab-scale manufacturing and upstream equipment, materials, and IP — are where the second mission's actual theory shows.
• TCS's Q1 FY27 earnings on July 9. The July 5 digest flagged it as "the first hard revenue read of the fiscal year" on whether AI-led pricing pressure is showing in reported numbers — the application-and-services layer's first quantitative test after a two-year, 46%-from-peak derating and the arrival of hyperscaler forward-deployed-engineering units aimed squarely at the SI layer's implementation revenue.
• Whether an Azure India region ever appears on the Claude-in-Foundry rollout, and whether the IndiaAI Mission publishes a utilisation record for its ₹65/hr compute. Both are the difference between an instrument announced and an instrument working — the residency gate actually opening, the subsidised-compute pipeline actually drawing down at scale.

What it buys

So: what does India's state-in-the-stack strategy actually buy?

It buys resilience and capital efficiency at the layers where those are achievable — packaging capacity on the mission's promised timeline, data-centre and subsea build-out that no foreign directive can switch off, subsidised compute and foundation-model support that compound slowly, and a governance perimeter drawn faster than most peer economies have managed. Those are real assets, and the week added concrete instances of each. The honest read is that India chose the layers where a non-frontier state's capital goes furthest, and it is executing that choice on schedule.

It buys nothing at the two layers the week showed to be state instruments: leading-edge silicon and frontier models. There, India is a taker, and June and July priced exactly what that costs — a nineteen-day worldwide shutoff, a US-only clearance tier, a government review step in front of general availability, and now the prospect of a shareholder-state behind the API. The strategy does not fix this exposure. It was never designed to; owning the frontier is the one thing India explicitly decided not to attempt.

What the strategy does — and this is the part the week clarified — is make the exposure survivable rather than fatal. The single most durable Indian response to June was not a policy or a subsidy but an architecture: keep a self-hostable open-weights fallback warm, because, as the July 2 digest put it, "an open-weights model on Indian GPUs cannot be switched off by a foreign directive." That sentence is the whole strategy compressed. India cannot own the frontier layer, so it is building the layers underneath it dense enough, and diversified enough, that when the frontier layer is used as an instrument against it, the fall is to a floor India controls rather than to zero. Whether that floor is high enough is the open question — it depends on the Indic model cohort closing the capability gap, on the ₹65/hr compute actually being drawn down, on ISM 2.0 becoming money. None of those is settled. But the shape of the bet is now fully legible, and the week of June 28 to July 5 is when it stopped being a policy preference and became a demonstrated necessity. India is not building the stack it would choose in a world where the frontier were neutral. It is building the stack it can afford in the world the frontier turned out to be.

Sources

• 2026-06-26. Anthropic Mythos 5 US-only clearance — CNBC, June 26, 2026 →; Axios, June 27, 2026. OpenAI GPT-5.6 preview — OpenAI, "Previewing GPT-5.6 Sol," June 26, 2026 →.
• 2026-06-28 (digest). India AI Digest 2026-06-28 — Washington clears Mythos 5; GPT-5.6 preview; the Indic-language column →.
• 2026-06-29. South Korea state-backed chip-and-AI drive — CNBC, June 29, 2026 →. Digest 2026-06-29 →.
• 2026-06-30. ISM 2.0 EFC clearance — Republic World / YourStory / Daily Pioneer, June 30, 2026 →. Claude Sonnet 5 — Anthropic, June 30, 2026 →. Digest 2026-06-30 →.
• 2026-07-01. Fable 5 restoration — Anthropic, "Redeploying Claude Fable 5," June 30, 2026 →; CNBC, June 30, 2026 →. RBI gap-assessment deadline — Business Standard, June 9, 2026 →. Digital India 11 years / ₹65-per-hour compute — ANI / MeitY, June 30, 2026 →. Digest 2026-07-01 →.
• 2026-07-02. Supreme Court on AI-hallucinated citations — Bar & Bench, July 2, 2026 →. I-2SEA subsea cable — Lightstorm, July 2, 2026 →. Digest 2026-07-02 →.
• 2026-07-02 / 07-04. OpenAI 5% US-government stake proposal and pre-release review framework — CNBC, July 2, 2026 →; Bloomberg, July 2, 2026 →; original reporting Financial Times. Digest 2026-07-04 →.
• 2026-07-03. MeitY separate-AI-legislation signal — Business Standard (PTI), July 3, 2026 →. Delhi HC Chadha deepfake order — SCC Online, July 2, 2026 →. Nasscom–Zinnov GCC report — press release via PRNewswire, July 2–3, 2026 →. Digest 2026-07-03 →.
• 2026-07-04. CG Semi Sanand OSAT inauguration — Akashvani News, July 4, 2026 →. Top-five IT market-value tally — Business Today, July 4, 2026 →. Anthropic offshore-access enforcement — Financial Times via BanklessTimes, July 3, 2026 →. Digest 2026-07-05 →.