The off-switch

Thematic essay — week of June 9–16, 2026.

On June 12, the US Commerce Department revoked a capability that paying Indian customers had been using for three days. Anthropic had launched Claude Fable 5 and the restricted Mythos 5 on June 9; on June 12 it suspended all access to both, under an export-control directive that covered "any location outside the US and all foreign persons inside it." Because the company could not separate foreign nationals from its user base in real time, the operative effect was, as the June 14 digest put it, "a hard global shutoff, not a geofence." India is Anthropic's second-largest market. The off-switch was not held in Bengaluru or New Delhi.

For two years, India's sovereign-AI argument was a principle in search of an incident. This was the incident. But the more useful thing the week did was hand India a diagnostic. If a foreign government can switch off a model your production stack depends on, the right question is no longer "is sovereignty worth it" — it is "which layers of this stack do we actually control?" The rest of the week, read in that light, is an audit. At each layer — frontier models, compute, domestic models, hardware, payment rails, the services channel, the law — the same question applies: does India own this, rent it, or merely host it?

The event: what an off-switch actually is

The mechanism is the part worth slowing down on, because the mechanism is what generalizes. This was not a price change, a regional licensing tier, or a capacity constraint. It was a government directive that revoked already-deployed capability from paying customers overnight, with no consultation and no transition window. The trigger is contested — per Axios and Fortune, the administration moved after a company claimed it had jailbroken the publicly available Fable 5 to surface cyberattack-relevant information, and Anthropic has characterized the finding as a misunderstanding. The trigger will be litigated; the precedent will not. As the June 14 digest framed the second of its two readings, this was "the first hard application of export controls to a deployed model's inference, not its weights — a precedent that converts model access into an instrument of trade policy."

That distinction — inference, not weights — is the whole lesson. A model you call over an API is not a model you have. The weights sit on someone else's servers, under someone else's jurisdiction, and the capability reaches you only as long as two governments and one company all keep the channel open. The June 14 digest reduced it to a sentence builders spent the weekend re-reading: "the variable that just changed is not cost or latency. It is whether the model is yours to depend on at all." Or, more bluntly: "frontier-model access via a foreign API is a privilege, not an asset on the balance sheet."

This is the top layer of the audit, and the verdict on it is the easiest to write. The frontier — Claude, GPT, Gemini — is rented. India's most capable AI capability is leased from US labs on terms a US agency can rewrite on a Friday. The same June 10 digest that covered Fable 5's launch had already noted the standing version of this constraint for regulated sectors: "the residency wall holds as ever" for BFSI and healthcare under DPDP, because "a US-hosted frontier model with no India-region inference stays out of regulated, data-resident workloads." June 12 extended that wall from a compliance problem to an availability problem. The specific models will likely come back. As the digest noted, "the knowledge that they can vanish on a Friday will not."

The audit: host, rent, own

Run the rest of the week's events through the same three-way sort and the Indian stack's real shape comes into focus. The distinction that does the work is between hosting something, renting it, and owning it — because the week supplied a clean example of each, and they are routinely conflated in the "India is building AI infrastructure" headline.

Take the layers in turn, because the interesting findings are at the seams.

Compute is mostly hosted, not owned. The June 10 digest's Meta–Reliance deal is the cleanest illustration in the archive of the difference. Reliance will "build and operate a 168 MW AI data center in Jamnagar, Gujarat, which Meta will lease," running on Indian renewables and desalinated seawater. Capacity is going up on Indian soil. And, as the digest was careful to state, "it is reserved for Meta's workloads … not for Indian builders to rent GPU hours against." The verdict it reached is the verdict the whole week argues for: "This is not India gaining sovereign AI compute. The capacity is leased to a US company for its own use … It adds to the data centers physically in India; it does not add to the compute Indian builders can rent." Hosting is not owning. A data centre you built for someone else, on land you own, powered by your grid, is still their compute. If a future export directive named compute the way June 12 named a model, the physical location of the building in Gujarat would not help the tenant.

The supply-side finance under that compute is offshore in the same way. The June 11 digest covered Apollo leading "an initial $35 billion capital solution for Broadcom's new AI XPV Platform," targeting "more than 20GW of compute capacity for frontier AI labs through 2028." The scale comparison the digest drew is the one to carry: that 20 GW "is roughly four times the largest announced India data-centre pipeline, financed in a single private platform without a state actor in the stack." The structural read followed — "the global compute divide is increasingly a private-capital-markets divide." India's policy instruments address the demand side of domestic compute, GPU empanelment and subsidized access; they do not touch the supply-side machinery where contracted frontier demand pulls tens of billions in private credit. That machinery is rented too.

The model layer is the one India is actively trying to own — and the week was its best in a while. Two events. First, the June 12 launch of Avataar's Varya, covered in the June 14 digest: "a 14-billion-parameter text-to-video model that generates a five-second clip in roughly 45 seconds on a single Nvidia H200," built under the IndiaAI Mission on subsidized national GPU capacity, with inference costed at "₹0.48 per second of video." The digest named exactly why it mattered in the week it landed: "the same week a foreign frontier model proved revocable, an IndiaAI-Mission startup shipped a model that India controls end to end." It also supplied the honest ceiling: "a 14B video model is not evidence India can match frontier scale across the board. It is evidence the cost-optimized, India-specific niche is buildable here now." Owned, narrow, real.

Second, the capital under the model layer. The June 15 and June 16 digests covered Sarvam closing "$234 million as the first tranche of a planned $300 million Series B," at roughly $1.5 billion — India's newest AI unicorn. The structural detail is not the unicorn label but the lead: HCLTech committed $150 million and led the round. As the June 15 digest read it, "an Indian systems integrator committing $150 million and leading the round is not passive growth capital," because it gives "an Indian SI an equity position in the model layer rather than a reselling relationship on top of it." That is an ownership move at exactly the layer the week proved was rented everywhere else. An SI that owns equity in a domestic model has a fallback that an SI reselling frontier APIs does not.

Both events carry the same caveat, and the chronicler's register requires stating it rather than burying it: a unicorn valuation is investor conviction, not a benchmark, and Varya's 10× cost-efficiency claim is Avataar's own, not independently reproduced. The June 16 digest's watch item is the right test — "a new Sarvam model release with published Indic benchmarks … is the test of whether this capital converts to measured capability." Owned-but-unproven is a better position than rented-and-revocable, but it is not the same as owned-and-competitive.

Hardware is owned at the design layer and aspirational below it. The June 11 digest's Zoho Nathu La is the precise version of this. Zoho's server platform runs on Intel Xeon silicon, so, as the digest said plainly, "this is not chip sovereignty." What Zoho owns is "the layer above the chip: board design, system integration, the patents on both, and the manufacturing relationships," with "all IP … owned in India." That is a real, narrow ownership claim — a layer "no Indian operator has owned at production scale before" — sitting on top of a chip layer that remains imported. The June 15 digest's Adani–Jabil announcement reaches for the layer below: an "intent to form a strategic alliance" to manufacture AI data-centre hardware in India, "making the hardware the build-out consumes rather than building the build-out," an "import-substitution play on AI data-centre hardware." Gautam Adani's framing was the week's most explicit sovereignty pitch — that the alliance would "ensure India is not merely a consumer in the AI age, but a creator, builder, and exporter of intelligence." The digest's verdict was equally explicit: "there is no JV, no facility, and no production line behind it yet … What exists today is a joint statement of intent." Aspiring to own is not owning.

The one layer India owns outright is the rails. The June 16 digest's Pine Labs P3P — "India's first agentic payment protocol," letting AI agents complete UPI payments after a single upfront authorization — sits on infrastructure no foreign agency can switch off, because UPI is public Indian infrastructure operated by NPCI under RBI. "Agent-initiated payments on UPI's low-cost rails open a distinctively Indian unit-economics path," the digest noted, and the constraint it surfaced is domestic, not geopolitical: "NPCI and RBI have not ruled on agent-initiated mandates. P3P ships into that gap." This is the inverse of the Fable 5 situation. The capability is early and the rulebook is unwritten, but the rails belong to India. A foreign export directive has no purchase on a UPI mandate. The contrast with the top of the stack is the essay's sharpest single point: India rents its smartest models and owns its payment layer, and on June 12 the second of those facts was worth more than it had been on June 11.

The services channel hedged. The June 13 digest's TCS–Anthropic partnership — provisioning Claude to "50,000 employees across 56 countries," the "first bilateral, named frontier-lab alliance for any of the Indian majors" — looks like deepening dependence, and partly is. But read against the same digest's note that TCS already runs Microsoft 365 Copilot at over 100,000 internal seats, the move is optionality: "Model-vendor optionality at the SI layer is now explicit TCS posture, on paper twice over." The Indian majors are not choosing a model; they are refusing to be locked to one. HCLTech's Sarvam equity is the most aggressive version of the same instinct — reach for the model layer directly. LTIMindtree's BlueVerse Currency, from the June 11 digest, hedges a different exposure: a pricing construct for the agentic era, because effort-based billing "prices the input that agentic AI is designed to reduce." Optionality is not ownership, but for a services book it may be the more rational posture — the SIs' product is the integration, not the model, and a vendor they can swap is a vendor that cannot switch them off.

What this looks like elsewhere

The instinct to read June 12 as uniquely Indian misses that the off-switch is a category the rest of the world has already met, and the comparison sharpens what India's actual position is.

China is the obvious case, and the instructive one. Export controls on advanced GPUs — not inference, but the chips underneath — forced a self-sufficiency program that produced DeepSeek: state-adjacent capital, open weights, aggressive pricing, frontier-adjacent capability shipped on time. The relevant difference is sequencing. China was cut off at the hardware layer years before it had a frontier model, and built the model layer under that constraint. India was cut off at the inference layer this week, after a decade of optimizing its economy around renting that layer. China's self-sufficiency is a finished argument with shipped artifacts; India's is, in the June 14 digest's exact phrase, still the point where "a discourse shift is not a capability shift."

The European comparison runs the other way and is the cautionary one. The EU has pursued sovereign cloud and a sovereign model champion for years — the regulatory apparatus is mature, the political will is real, and the capability gap to the US frontier has nonetheless not closed. Sovereignty as procurement preference and statute, absent a shipped frontier model, buys residency and optionality but not capability. That is roughly the position the June 11 digest's Vaishnaw item describes for India: a minister signaling that "the world of AI is very different from the world when the IT Act was enacted in 2000," a purpose-built statute coming, but — as the digest noted — "what it does not signal is what the statute contains." A framework is forming; its perimeter is unknown. Europe is the evidence that the framework is necessary and nowhere near sufficient.

Where India has a genuinely differentiated card is the one the audit surfaced: the public-rails layer. No other large economy has a state-owned, population-scale digital payments and identity stack of the India Stack's reach to build agentic AI on top of. Sarvam co-founder Vivek Raghavan made exactly this argument at the February summit, leaning on his Aadhaar and India Stack experience to frame the choice, in the June 14 digest's quote, "between sovereignty and becoming a 'digital colony.'" The Pine Labs protocol is the first concrete instance of AI being wired into that owned layer. The comparable is not Europe's sovereign cloud or China's chip program; it is that India's sovereignty edge, if it has one, is at the application-and-rails layer it already owns, not at the model-and-compute layer it is trying to buy its way into two orders of magnitude behind.

Where it lands

The audit produces watch items, not predictions, and the week attached a concrete trigger to each.

On the rented layer: the terms of return. Watch whether Fable 5 and Mythos 5 come back, and how. The June 14 digest drew the distinction that matters: "a quiet reinstatement and a permanent carve-out regime are very different signals." Watch, too, whether the next foreign frontier releases — Gemini 3.5 Pro's general availability and Grok 5 — "come with any India-region or residency commitments, or arrive on the same revocable terms." If they arrive revocable, the rented-layer verdict hardens and the case for the owned layer strengthens with it.

On policy: a circular, not a speech. The June 14 digest's test is the right one and worth quoting as the bar: "Statements are cheap this week; a procurement circular would be the real tell." The specific signal is whether MeitY or the IndiaAI Mission issues guidance on frontier-model dependency for government and regulated deployers, or pushes domestic-model procurement preference for state workloads. A discourse that wins the weekend converts to capability only if it shows up as a procurement line.

On the owned-but-unproven layer: a benchmark. The single most informative event over the next two quarters is a benchmarked release that tests an ownership claim — Sarvam's next model with published Indic benchmarks, or independent quality-and-cost numbers on Varya from outside Avataar. Owning a model the frontier can revoke around is worth little if the model is not good enough to fall back to. The benchmark is where owned-narrow either becomes owned-competitive or stays a press release.

On the rails: an NPCI ruling. The June 16 digest's watch item — "an NPCI circular or RBI guidance specifically addressing AI-agent-initiated UPI mandates" — is the signal that turns India's owned layer from a demo into deployable infrastructure. The rails are India's; whether agentic AI can run on them at scale depends on a domestic regulator, not a foreign one. That is the whole point of owning the layer.

On hardware: a signed line. Adani–Jabil converting "intent" into a definitive JV, a named site, or a groundbreaking is the test of whether the hardware-ownership ambition is real. Until one lands, the layer below Zoho's board design stays imported.

The honest answer

So: what does India own?

The audit's answer is more precise than either the boosterist or the dismissive read. India hosts a great deal — Jamnagar's megawatts, the hyperscaler builds, the GCC labor that the June 10 Opendoor item showed is itself now exposed to AI substitution rather than relocation. India rents the layer that matters most — the frontier models its SIs, startups, and regulated enterprises had quietly wired into production, and that a US agency switched off for three days to demonstrate it could. And India owns, outright, a thinner band than the headlines imply: domestic models that are real but sub-frontier, a server-design layer sitting on imported silicon, and — the one unambiguous, uncontested holding — the public payment and identity rails that no export directive can reach.

June 12 did not change any of those facts. It made them legible. Before the cutoff, "hosting," "renting," and "owning" all photographed as "India is building AI infrastructure." After it, the difference between the data centre Meta controls in Gujarat and the UPI rails India controls everywhere is not a rhetorical one — it is the difference between capability that can be revoked from Washington and capability that cannot. The sovereign-AI argument spent two years asserting that distinction mattered. The week supplied the demonstration, and then, across six other events, an inventory of exactly where India sits on the right side of the line and where it does not.

The gap at the model-and-compute layer did not close this week, and the numbers say it will not close by out-spending or out-computing a frontier financed in $35-billion private-credit tranches. But the week clarified where the durable Indian positions actually are: not in matching the frontier, but in the layers India can own outright — language, residency, and the public rails the rest of the stack runs on. The off-switch was a loss. The diagnostic it handed over may be worth more than the three days of access it cost.

Sources

• 2026-06-10 (digest). India AI Digest 2026-06-10 — Meta–Reliance 168 MW Jamnagar lease; Opendoor India exit; Claude Fable 5 / Mythos 5 launch and pricing →. Primary: TechCrunch (Meta–Reliance); TechCrunch (Opendoor); Anthropic (Fable 5 / Mythos 5).
• 2026-06-11 (digest). India AI Digest 2026-06-11 — Zoho Nathu La server platform; Vaishnaw on a new AI law; Apollo–Broadcom $35B / 20 GW XPV financing; LTIMindtree BlueVerse Currency →. Primary: Business Wire (Zoho); Inc42 (Vaishnaw); Apollo (Broadcom XPV).
• 2026-06-13 (digest). India AI Digest 2026-06-13 — TCS–Anthropic regulated-industries partnership →. Primary: Anthropic (TCS).
• 2026-06-14 (digest). India AI Digest 2026-06-14 — US export-control cutoff of Fable 5 / Mythos 5; Avataar Varya launch; the sovereign-AI posture shift →. Primary: Anthropic (access suspension); Axios; PIB (Varya).
• 2026-06-15 (digest). India AI Digest 2026-06-15 — Sarvam $234M Series B / HCLTech lead; Adani–Jabil hardware-manufacturing intent →. Primary: TechCrunch (Sarvam); Adani (Jabil alliance).
• 2026-06-16 (digest). India AI Digest 2026-06-16 — Pine Labs P3P agentic payments on UPI →. Primary: MediaNama (P3P).