Sovereign AI, Made Elsewhere

A small notice went out from Brasília on Friday and was made public yesterday. SERPRO, the federal technology company that holds the data spine of the Brazilian state, has chosen its partner for "IA Soberana", Brazil's national sovereign AI programme. According to the announcement carried by PR Newswire, the winner is MeetKai Brasil, the local arm of a Los Angeles company. Brazil's broader AI plan envisions roughly R$23 billion through 2028, large language models trained in Portuguese, all of it running on Brazilian infrastructure.

It is worth pausing on what sovereign means here. The vendor is foreign. The weights, the operating control, the data, the language of the model: all of these will sit inside Brazil. Sovereignty has been defined not as "we built it ourselves" but as "we hold the keys, in our language, on our soil". A country with a serious public data estate decided that the architecture of control matters more than the nationality of the builder.

For anyone watching India's own AI conversation closely, this is a useful distinction to import. Sovereign AI is not one thing. It is a stack with at least six layers: data, compute, foundation model weights, fine-tuning, hosting, governance. A government has to decide, layer by layer, which it must own, which it can lease, and which it should regulate without ever touching. Owning everything is expensive and slow. Owning nothing is a different kind of dependence. The hard work sits in the middle.

What India has done well, through a decade of digital public infrastructure, is own the rails: identity, payments, consent. What it has not yet decided, in public, is how much of the model layer above those rails should be Indian by ownership rather than Indian by use. The Brazilian tender is a quiet reminder that the answer need not be all or nothing. It can be this: we will host it, we will train it on our languages and our case files, and we will fire the vendor if they misbehave. That is a thinner sovereignty than the slogan, but it travels further.

The next time "sovereign AI" comes up in an Indian conference room, the right first question is not who built it. It is which layer of the stack you mean.

#SovereignAI #AIPolicy #PublicSector #IASoberana #DigitalSovereignty #IndiaAI #AIGovernance #DPI

AI Just Became Capital Risk

On 24 June, the Reserve Bank of India dropped a draft that does not look like an AI regulation at first read. It is titled Guidance on Regulatory Principles for Model Risk Management, 2026, and the language is pure prudential supervision: board-approved frameworks, three lines of defence, independent validation, risk tiering, model inventories retained for ten years after decommissioning. Spend twenty minutes with it and a sharper thought lands. India's central bank has just pulled AI into the family of supervisory tools normally reserved for capital and liquidity. The principle that a regulated entity remains fully accountable for the outcomes of every model it uses, including those sourced from third-party vendors, quietly closes a loophole that the entire fintech-banking AI stack had been leaning on.

What the draft actually says

The reach is wide. Eleven categories of regulated entities — commercial banks, small finance and payments banks, urban and rural co-operatives, NBFCs across all layers, AIFIs, ARCs, credit information companies — are inside the perimeter. The definition of “model” is wider still. It covers not just neural networks and generative AI but also algorithms, analytics tools, decision rules, even spreadsheet-based tools where they materially influence credit, pricing or risk decisions. Every such model must sit under a Board-approved Model Risk Management Framework. Every high-risk model must be cleared by the Risk Management Committee of the Board before deployment. Every customer-facing AI interface must disclose that it is AI, list its limitations, and offer the user a way to switch to a human. And every deployed AI system needs a working kill switch — a one-button deactivation if the outputs go wrong. Comments are invited until 24 July.

The deep idea: models as a new line on the risk ledger

This is the consequential part, and most early commentary is missing it. When a regulator says you cannot offload accountability to a vendor's API, it is doing for AI what Basel did for credit risk in the early 1990s. Models become a class of exposure. They have to be inventoried, tiered by materiality, validated independently, monitored for drift, and signed off at the highest level. That is the grammar of capital, not the grammar of code.

It is also the only grammar that scales. The RBI's own FREE-AI Committee survey last year found that nearly 21% of regulated entities were already deploying AI in production — across credit underwriting, cybersecurity, customer support, sales — and 67% wanted to go deeper. At that level of penetration, telling boards “your tech team has this” is not serious supervision. Every credit cycle eventually meets its model failure mode. A regulator that waits to discover the failure is a regulator already too late.

The kill switch rewrites the vendor market

Look at the supply side and the picture sharpens. A small set of global tech firms quietly supplies a disproportionate share of the AI models running inside Indian financial services. The draft flags this concentration explicitly as a systemic supply-chain risk. Combine it with the third-party-is-no-defence clause and the direction of travel is unmistakable: a serious push toward indigenous, auditable, swap-out-able model stacks. The new market is not the bank's market. It is a RegTech market — model validation firms, bias-testing labs, explainability auditors, red-teaming shops. A compliance officer reading this draft on Monday morning is realising that the next critical hire is not another data scientist; it is an independent validator who can sign off on someone else's data scientist.

Why this logic will travel beyond banking

I have spent enough time inside a national administration deploying AI on citizen-facing systems to read this draft as a template, not an end-point. Tax, customs, social security, urban service delivery, public health records — every government body running models on millions of files faces the same accountability question the RBI is now putting to banks. Who signs off on the model that decides a refund? A scrutiny notice? An eligibility threshold? The answer cannot be “the algorithm”. It has to be a named human, with a department behind them, with an institutional framework above them.

Three things from this draft will, I think, become the standard public-sector discipline within two budget cycles:

• the inventory rule — every active and decommissioned model on a register, with a ten-year tail;
• the explainability threshold — outputs interpretable to the extent the business process actually requires;
• the kill switch as a non-negotiable product feature, not a nice-to-have.

The trade-off, and the right side of it

The cost is real. Industry analysts already estimate a 50 to 100 basis point rise in IT spending for serious adopters. Smaller NBFCs and co-operative banks will feel it the most. Time-to-market for AI-driven products will lengthen. None of this is free, and the smaller institutions will need a transition window the draft does not yet promise.

But picture the alternative. A black-box model, bought from a vendor, denying loans to a cluster of small borrowers in a particular district, operating under no one's clear authority. Someone eventually discovers it — they always do. The regulator pays, the bank pays, the customer pays, and trust in AI itself pays the heaviest tax of all. The cost of the RBI's draft is the cost of preventing that discovery.

The cleaner way to say it: in a regulated industry, AI stops being a technology decision and becomes a capital decision. Capital decisions are made in boardrooms, not in model repositories. The window closes on 24 July. The more interesting question is not whether this framework is right for finance — it broadly is — but how quickly the rest of the public-private edge of AI adoption catches up. The RBI has, almost without saying so, written the first chapter of an Indian AI accountability code. Other regulators will write the rest, or they will inherit the failures of not having done so.

#RBI #AIGovernance #ModelRisk #BankingRegulation #IndianEconomy #AIAccountability #PublicFinance #RegTech

Korea Was The Canary

On Tuesday, June 23, South Korea's KOSPI fell 9.99 percent in a single session, tripping circuit breakers, wiping out roughly $2.5 billion in foreign capital in hours, and qualifying as the fifth-largest single-day decline in the index's history. Samsung and SK Hynix each lost more than twelve percent. The Nasdaq followed down 2.21 percent the next session. Oracle, in the same news cycle, disclosed it had cut twenty-one thousand jobs in a year — almost thirteen percent of its workforce — and named AI as the reason. Forty-eight hours, three disclosures. None of this is a tech story. It is the beginning of a macro story we have not yet learned to read.

The capex has become the commodity

For two decades we taught ourselves that crude was the single variable that synchronised global cycles. A spike in oil touched everything: inflation in importers, fiscal space in exporters, central bank reaction functions everywhere. That intuition is still half right. But a new variable has joined it, and the last week suggests it is, in the short run, more potent.

Meta, Google, Microsoft, Amazon and Oracle are between them committing capex plans this year that could touch seven hundred billion dollars to build AI data centres. Oracle alone reported negative free cash flow of $23.7 billion last fiscal year while raising capex 162 percent to $55.7 billion. Those numbers are not technology numbers anymore. They are macroeconomic numbers — comparable in scale to the annual oil import bills of mid-sized economies — and they are decided in a handful of US boardrooms.

This is the kind of single-factor dependence Professor Richard Robb's International Capital Markets course at Columbia kept circling: when cross-border flows are tethered to a small set of decisions on a small set of US balance sheets, the receiving economies inherit volatility they did not choose and cannot hedge. Korea just lived through one rehearsal.

Why Korea fell first

Korea was not a random victim. The KOSPI was up roughly 95 percent year-to-date going into Tuesday. Samsung and SK Hynix together account for about half the index by market capitalisation. The Bank of Korea has openly said AI-related chip exports will add 0.7 percentage points to 2026 growth, more than offsetting the drag from costlier oil. Taiwan is on track for 9.6 percent GDP growth this year — its highest in sixteen — on the same trade.

When the global market began doubting whether US hyperscaler capex was sustainable, every one of those exposures got marked at once. Three triggers converged on the same morning: MSCI again excluded Korea from its developed-markets watchlist, regulators raised flags about leveraged single-stock ETFs tied to Samsung and SK Hynix, and a hawkish Federal Reserve dot-plot from June 17 was already in the bloodstream. The market did not need a new fact. It needed a coordination point.

India's awkward middle position

India's place in this story is uncomfortable. Unlike Korea or Taiwan, India is not a meaningful seller into the AI hardware stack. Unlike China, it is not building frontier models at scale. The result is the worst of both worlds: when AI capex booms, India captures little of the upside; when it wobbles, the contagion still arrives — through portfolio outflows, currency pressure and the generic risk-off impulse against emerging markets.

The numbers this month are blunt. Foreign portfolio investors pulled roughly sixty-four thousand crore rupees out of Indian equities in the first half of June alone, the heaviest exit since March, with elevated oil and "concerns over AI's impact on tech revenues" cited as the principal reasons. Two macro factors, neither of which India controls, set the direction of an enormous slice of market cap. That is not market accident; that is structural exposure.

What policymakers should actually do

One. AI capex belongs on the macroprudential dashboard. The Reserve Bank's Financial Stability Report already tracks crude, dollar moves, FII positioning and banking-sector stress. It should now also track the announced capex plans of the five US hyperscalers, because in any given quarter those plans are a bigger swing factor for emerging Asia than the OPEC+ communique. Treating this as a tech-sector story is a category error.

Two. The export-services tax base — IT, ITES, global capability centres — is more cyclically exposed to AI capex than its standard sector classification implies. Revenue projections and advance-tax assumptions should stress-test against a fifteen to twenty percent compression in this base, not as a tail risk but as a plausible scenario for the coming eighteen months. A tax administration cannot afford to be the last institution to learn that a sector's cycle has changed.

Three. The Indian debate around "missing the AI boat" oscillates uselessly between buying chips and drafting strategies. The better path runs through demand the country actually controls — large-scale public-sector AI deployment in tax, courts, health, urban services — so that compute spend, even if imported, gets monetised at home through productivity. A country that is a net buyer of AI inputs must, at minimum, be the most efficient internal consumer of them.

The bigger lesson

The KOSPI's nine-point-ninety-nine percent is not really a market story. It is a structural disclosure. Forty years ago a single oil price ran the world's inflation and growth narrative. We are not there yet with AI capex. But we are closer than is comfortable, and the trajectory is one-way. The job of policymakers in countries that neither make the chips nor own the models is to stop treating each AI-driven wobble as a curiosity and start treating it as a recurring macro shock with the same seriousness we reserve for crude.

Korea was the canary. The mine is the rest of us.

#AIcapex #KOSPI #EmergingMarkets #IndiaEconomy #GlobalMarkets #Semiconductors #MacroPolicy #ForeignFlows

Government Isn't A Hyperscaler

On Monday, Alphabet fell about 5%, dragging the communication services sector down with it. Memory chip stocks plunged in Asia overnight and the selling crossed the Pacific by Tuesday morning. Beneath the price action, the cause was specific and revealing: investors are starting to ask whether the enormous sums being poured into artificial intelligence will earn their keep, and reports of senior talent leaving Alphabet's AI teams added to the unease.

That question — will the AI capex pay off — is the right question for shareholders. It is the wrong question for a government.

What the market wobble actually says

The sell-off was narrow, not broad. The Russell 2000 closed above 3,000 for the first time even as the big tech names tumbled. That tells you the market is rotating, not collapsing. The doubt is concentrated in a handful of names whose valuations had assumed the AI trade would compound for several more years without a pause.

The question being asked is simple. Amazon, Microsoft, Alphabet and Meta have collectively guided to roughly 25 billion in capital expenditure for 2026. Their combined free cash flow is forecast to fall to about billion in the third quarter — a decade low. At that ratio, the cash being spent is no longer comfortably backed by the cash being earned. That is what the nervous side of the market is pricing.

Why this question doesn't translate to government

Public administration does not sell tokens. It does not have shareholders. Its “revenue” is taxes and its “product” is rights, services, and the slow accretion of public trust. The whole logic of equity-style ROI — free cash flow divided by capex — is structurally absent.

And yet, watch any AI tender written by any government this year and you will see private-sector language smuggled in. Pilots. Use cases. Productivity gains. Cost savings per FTE. We are evaluating public AI investments in the only vocabulary the consultants brought to the room, and the result is that everyone ends up arguing over a metric nobody in the citizen-service chain actually cares about.

From inside a national tax administration, having worked on conversational AI for citizens at scale, I can say something specific. The right number was never “queries handled per rupee of compute.” It was something closer to: did the next confused taxpayer get an accurate answer, in the language she actually speaks, in time to act on it, without having to travel to a counter? That is an outcome question. It is not a capex question.

An outcomes-first lens

If the market is pricing AI on cash flow, the government should price AI on three different things:

• Service latency. How much faster does a citizen get an answer, a refund, or a decision after the system is deployed?
• Service reach. How many more citizens, in how many more languages and pin codes, can the same service touch without adding counters?
• Officer leverage. How much higher-order work can the same officer do because the system has absorbed the routine?

None of these are radical. They are simply restatements of what the public sector exists to do. The discipline is to write them into the procurement document before the vendor walks in, not after the dashboard is built.

A proposal: separate the two stacks

Here is a concrete proposal. A government should split its AI estate into two stacks and evaluate them by two completely different rules.

The internal stack — drafting, summarisation, case retrieval, file review, internal search across decades of orders — can be evaluated commercially. Hours saved, error rates, contractor-substitution ratios. That is fair, because the work being replaced was already priced.

The citizen-facing stack — multilingual chatbots, eligibility navigation, guidance on a new statute, status tracking on a refund — must be evaluated as service delivery. The metric is not unit economics. The metric is whether a person previously locked out of a service is now inside it.

Confuse the two stacks and the second one will always lose to the first on a finance committee's spreadsheet. It will be quietly defunded the moment an AI valuation cycle turns. And, as Monday's tape made plain, the AI valuation cycle will turn.

The public sector organisations point

Professor Michael Ting's course on the Analysis of Public Sector Organizations at SIPA made an argument that has aged into something close to a law. Public agencies serve multiple principals — legislatures, ministers, courts, auditors, citizens — and the metrics they adopt silently decide which principal they end up serving. Adopt private-sector ROI metrics for public AI and the system will end up optimising for the auditor's spreadsheet, not the citizen at the window. That is not a hypothetical risk. It is the default trajectory.

What to do this quarter

Three moves, for any department about to sign an AI contract:

• Write the outcome metrics before the price negotiation, not after.
• Build at least one citizen-facing metric whose headline a non-technical minister can defend in question hour.
• Keep funding the citizen stack even when the headline AI trade in the markets is going through one of its periodic doubts. Especially then.

The market is reminding everyone this week that the private bet on AI is a bet. Government is not making the same bet. It is deploying a tool that should be judged by whether the queue moved. The discipline is to keep saying that out loud while the news cycle is busy saying everything else.

#AIinGovernment #PublicSectorAI #DigitalGovernance #AICapex #IndiaAI #Governance #PublicFinance

The Filing India Doesn't Owe

On 30 June 2026, exactly a week from now, multinational groups across more than three dozen jurisdictions will lodge their first GloBE Information Return. It is the most ambitious cross-border tax filing the world has ever attempted. India is not in the queue.

That is not a footnote. It is the single most important fact about Indian international tax this year, and almost nobody is talking about it.

The week the global minimum tax goes live

The OECD's Pillar Two is now operational. The rules apply to multinational groups with consolidated revenues of at least €750 million. If their effective tax rate in any jurisdiction falls below 15%, someone, somewhere, collects a top-up. Roughly 140 jurisdictions joined the inclusive framework back in 2021. Thirty-seven have actually legislated a Qualified Income Inclusion Rule or a Qualified Domestic Minimum Top-up Tax that bites from the 2024 reporting fiscal year.

The OECD was still soldering plumbing in the last weeks before the deadline. On 18 May 2026 it issued a common understanding allowing groups to file centrally in one jurisdiction and avoid duplicate local returns, provided each domestic office gets a notification. That a structural feature of the regime had to be settled eight weeks before the first deadline tells you something about how unfinished this compact still is.

Three countries that stayed away

The United States, China and India have not implemented Pillar Two. The American position is now formal: in January 2026 the US Treasury announced that US-headquartered groups would be exempt, and the OECD's 5 January Side-by-Side package legalised that exit by carving out a safe harbour for groups parented in jurisdictions with ‘eligible’ tax regimes. The US is, as of today, the only jurisdiction on the OECD Central Record with a confirmed Eligible SbS Regime.

India's stance has been quieter, and to my mind more considered. We participated in the framework. We never legislated the rules. We watched.

Why I think the wait was right

Pillar Two is not really a tax. It is a coordination mechanism that exports one country's view of what another country's tax base should be. It is the first time in modern fiscal history that a domestic legislature has been asked to outsource the residual taxing right on profits earned at home to a residence jurisdiction abroad. That deserves more scepticism than it gets in polite international tax conversation.

For India, the cost-benefit was always thin. The corporate rate here is 22%, or 25.17% with surcharge and cess. The concessional rate for new manufacturers is 15% statutory, around 17.16% effective. We are not a low-tax jurisdiction. Pillar Two only matters when you are below the floor; we mostly are not.

The price of joining would have been real. A QDMTT regime to design. A GloBE return architecture to build. A workforce to retrain in jurisdictional ETR computation. Disputes to defend under accounting standards that are not ours. And acceptance that the rules will be rewritten by an OECD working party in which our vote is one among many. None of that grows the base. It expensively confirms what we already collect.

What we still need to claim

The harder side is data. The 2024 GIR is the first time multinational groups will publish, in a standardised XML schema, a jurisdiction-by-jurisdiction picture of where their profits arose and what tax those profits paid. Other administrations will receive that file automatically. India, outside the exchange relationships, will not, unless we sign on to receive it. That is a transparency dividend we should be claiming whether or not we ever impose a single rupee of top-up tax.

Soft power is the other piece. The Side-by-Side carve-out is, in effect, a US-only privilege today. The OECD has signalled other jurisdictions may be added. ‘May’ is doing the heavy lifting in that sentence. If a future investor's post-tax compliance burden is lighter under a US parent than under an Indian one, we have not lost any revenue, but we have lost a piece of the architecture of who matters in the next decade of international tax rule-writing.

A proposal

The smart move is not to copy Pillar Two. It is to ask for the data, build the analytical capacity, and use the next eighteen months to find out where Indian profit shifting is genuinely costing us revenue. Three concrete steps.

• Sign the GIR Multilateral Competent Authority Agreement. The cost is administrative. The value is a structured view of every in-scope group's worldwide tax footprint, delivered automatically.
• Commission a quiet domestic study of jurisdictional ETRs for India-headquartered MNEs. Two or three years of clean data will tell us whether a future QDMTT would collect ten thousand crore or ten lakh. Right now the number is asserted, not measured.
• Use the Income-Tax Act 2025 transition window to bake in a minimum-tax-compatible computational backbone. If we later choose to switch on a QDMTT, the systems already speak the schema. Building it after a political decision is far more painful than building it now, while the law is still warm.

Pillar Two will either be remembered as the most consequential multilateral tax instrument since the League of Nations drafted the first model treaties in the 1920s, or as a noble experiment that fractured the moment its largest economy walked away. We do not need to guess which today. We need to stay liquid: positioned to step in, positioned to step out, owning the data either way. That is the case for sitting out 30 June with intent.

#PillarTwo #GlobalMinimumTax #IncomeTax #IndianEconomy #TaxPolicy #OECD #InternationalTax

When Ships Paid In Renminbi

Somewhere in the European Central Bank's annual report on the international role of the euro, published a few days back, there is a sentence that ought to have set off more alarms than it did. Settlement activity on China's Cross-Border Interbank Payment System rose by more than a third in the days around the outbreak of the Middle East war. Quieter still, the report notes that some ships made payments in renminbi via CIPS, or in crypto-assets, to transit through the Strait of Hormuz during March and April. A handful of transactions, in absolute terms. But a tell.

The dollar is not dying. It does not need to die for the world's payment plumbing to start looking less like a single pipe and more like a switchboard.

The Hormuz Tell

For seventy years, the working assumption of treasurers and central bankers has been that when a regional crisis flares, dollars buy you out of trouble. They buy fuel, they buy insurance, they buy the wire that gets a cargo moving. The ECB's detail tells us that assumption is starting to fray at the edges. When a ship in a hurry could not, for whatever reason, settle a passage in dollars, the alternative chosen was renminbi or, more telling, crypto.

The size of the shift in flow is worth dwelling on. Customer-related cross-border payments by Chinese banks in renminbi reportedly hit USD 1.4 trillion in March 2026, roughly 30% higher than the previous month. CIPS itself had grown a meagre 3% in 2025, after years of 20%-plus expansion; the war reversed that deceleration in a single month. None of this dethrones the dollar. All of it builds optionality for whoever pays attention.

A Defensive Pile Of Gold, And Not Much Else

India sits in this picture awkwardly. We have, by ECB's reckoning, added about 130 tonnes of gold to the reserves stack since 2022, alongside Turkey, China and Poland. That is a sensible defensive instinct after watching Russia's reserves get frozen. It is also where the imagination seems to have stopped.

The position is genuinely odd. Of the major economies thinking about reducing dollar dependence, we are the only one that already runs a domestic retail payments network most of the world envies, settles more real-time transactions than the rest of the planet combined, and has spent a decade quietly exporting that stack to friendlier jurisdictions. From inside a national administration that has built and operated citizen-facing digital infrastructure at scale, I can say with some confidence that the hard engineering problem of cross-border instant payments was solved at home long before it became geopolitically interesting. What is missing is the political will to treat the stack as strategic infrastructure rather than a soft-power side project.

Wrong Question, Right Answer

Most Indian commentary on the de-dollarisation theme falls into a trap: which bloc should we join, the dollar one or the renminbi one. This is the wrong question, and it is asked by people who confuse a payment rail with a treaty. India's interest is not in joining a bloc. It is in being able to settle, invoice, hold and route in whichever instrument is cheapest, safest and least politically charged on the day a particular contract has to close. The word for that is optionality, and it is built, not declared.

Three moves are worth making.

Treat payment corridors as foreign policy

UPI acceptance in the Gulf, Southeast Asia, parts of Africa and small island economies is being built piecemeal, often as tourism convenience. Stop calling it that. A corridor that lets an Indian importer pay a Vietnamese supplier in rupees or dong, without a dollar leg, is strategic plumbing. Build it with line items in the budget and missions actively negotiating acceptance, the way other countries negotiate visa-on-arrival regimes.

Insist on settlement clauses in trade agreements

The next ten free trade agreements India signs should, at minimum, contain a clause allowing settlement in either party's currency for a defined share of trade, with central bank windows providing convertibility at agreed bands. Project mBridge, the multi-CBDC platform connecting China, Hong Kong, Thailand, the UAE and Saudi Arabia, shows what a serious version of this looks like. India is conspicuously absent from that table; that is a choice we keep making by default.

Stop confusing gold with a strategy

A 130-tonne pile is a backstop. It is not active monetary architecture. The state can hedge tail risks in metal and at the same time build the live system that determines who pays whom in normal times. The two are not substitutes.

What The Classroom Did Get Right

Years ago at Columbia, in Prof. Richard Robb's International Capital Markets class, the formative lesson was that the international monetary system runs on inertia. It does not change because a paper is published or a summit is held. It changes when, in some operationally tedious moment a ship in a hurry, a bank under sanction, a captain refusing to wait the cheaper, safer instrument turns out not to be the dollar. That moment does not need to come at scale to matter. It needs to come reliably enough that the next CFO writes the alternative into her treasury policy as a permanent option, not an emergency one.

Xi Jinping's 1 February call for the renminbi to become a global reserve currency, taken alongside CIPS opening up to multicurrency settlement from the same date, is best read in that light. It is not a slogan. It is a procurement plan for whichever country wants to underwrite the next system.

The Habit, Not The Asset

The dollar's strength has never been an asset class. It is a habit, a default keystroke on every treasurer's terminal. Habits are sticky. They are also vulnerable to small, repeated, observable counter-examples. A few ships at Hormuz paying in renminbi, or in stablecoins routed through some Gulf trading hub, will not by themselves break that habit. But anyone reading central bank reports for a living should treat what happened in March and April as the rehearsal it was.

India holds the world's best retail payments stack and a serious geopolitical hand. Both are pointed away from each other today. The most useful thing the Government of India could do this year is to draw a line between them: to convert a fintech achievement into an instrument of statecraft. The next time ships are in a hurry at a chokepoint, the question worth being able to answer is whether any of them are settling in rupees.

#InternationalFinance #DollarDominance #UPI #Renminbi #CIPS #Geoeconomics #IndianEconomy #PaymentSystems

When GeM Learns Tamil

A handloom weaver in Karur. A carpenter in Bhubaneswar. A metal-fabricator in Rohtak. None of them has ever sold to a government department through GeM. Not because their work is poor; because the tender document is in English.

Last week, on 15 June, the Digital India BHASHINI Division and the Government e-Marketplace signed an MoU to integrate AI-powered language technologies into GeM, letting buyers and sellers transact across the platform in 22 Indian languages, with voice as a first-class input. The press release was polite and technocratic. Underneath it sits one of the more consequential pivots in Indian e-governance in years.

What was actually signed

GeM is the national procurement portal under the Ministry of Commerce. Government buyers, from ministries down to municipal corporations, use it to purchase goods and services from registered sellers. The transaction volumes are not small. Until last week, the platform spoke fluent English with Hindi as a polite afterthought. The new MoU plumbs the marketplace with BHASHINI's translation APIs, voice bots, domain-specific language models, and a voice-first interaction layer.

This is not a one-off. In recent weeks BHASHINI has signed similar agreements with DPIIT for the industrial ecosystem, with the Ayush Ministry for traditional health knowledge, with the Centre for Railway Information Systems, and with Kathmandu University to extend Indian-language AI into South Asia. The portfolio matters more than any single MoU. Together, they amount to a slow, deliberate plumbing of the Indian state with a shared language layer.

Language is infrastructure, not a feature

Many departments still treat language support as a checkbox. Get the portal translated to Hindi, drop in a few Tamil PDFs, declare victory in the next presentation. That mental model is wrong. Language is not paint applied at the end of a project. It is the medium in which citizens think, decide and act. A self-employed entrepreneur in Madurai will not navigate a fourteen-page English tender to bid on a six-lakh contract, even if a translation button exists somewhere.

What BHASHINI is building is something different: a horizontal language layer of the Indian state. The platform supports 36 Indian languages for text translation, 23 for voice recognition and speech synthesis, and reportedly processes more than 15 million AI inferences a day. Once that layer exists, any new government service, from a procurement portal to a hospital appointment system to a tax helpline, can become multilingual at the speed of an API call rather than at the speed of a fresh translation tender.

This is how digital public infrastructure should be built. UPI did it for payments. Aadhaar did it for identity. DigiLocker did it for documents. Language was the last big primitive that every department was hand-rolling badly for itself. It now has a shared standard.

The voice part, which most commentary will miss

The MoU foregrounds voice. Translation is the obvious win; voice is the deeper one. A welder in Latur with a smartphone and modest literacy will not type a procurement query into a form. He will speak it. Voice agents that understand Marathi, Telugu, Bengali and Bhojpuri properly are how the next two hundred million Indians will actually interact with the state.

From inside an administration that serves taxpayers at population scale, I can say this with some conviction: the bottleneck is rarely the legal text. It is the gap between the citizen's spoken question and the legal text's English answer. Closing that gap, in real time, across 22 languages, is harder than building any specific service. It is also more transformative.

Platform thinking, not product thinking

I have argued before, and still believe, that public-sector AI strategy which begins and ends with “let us build a chatbot” misses the point. A chatbot is a user interface. An interface alone solves nothing if the underlying service, data and language plumbing is broken. The BHASHINI-GeM partnership matters precisely because it inverts the usual order. It does not ship a flashy procurement assistant. It plumbs the marketplace with multilingual capability so that any future interface, whether chat, voice, IVR or mobile app, can speak the citizen's language without each department rebuilding the linguistics from scratch. That is platform thinking, and it is the right thinking for a state with thousands of services and finite budgets.

Three things that must follow

If this is to scale beyond a press release, three things should be insisted upon. One, every central ministry should publish a short, public roadmap for plugging at least one citizen-facing service into BHASHINI within six months; pick the highest-volume interaction and start there. Two, language quality cannot be self-certified by the platform that builds it; an external panel of native speakers, including from non-scheduled languages, should publish quarterly benchmarks the way RBI publishes inflation prints. Three, the income tax, GST and welfare-delivery systems, which together touch nearly every Indian, must move early. That is where the political payoff is largest, and where the harm of getting it wrong is also greatest.

Two final observations. The Eighth Schedule lists 22 languages, but India actually speaks well over a hundred; treating BHASHINI as the floor, not the ceiling, will matter. And this kind of infrastructure rarely makes news on the day it launches. It shows up two years later, when a self-employed seamstress in Tirupur wins her first government order in Tamil and does not remember a time it was otherwise.

#BHASHINI #GeM #DigitalIndia #AIGovernance #PublicProcurement #MultilingualAI #VoiceFirst #DPI