The next 3 years with AI integration

Sovereign stacks, rented compute

Nations are pulling money, energy, data and AI inside their own borders. Most will still rent the hardest layers from two or three suppliers. This brief checks that thesis against where things stand today, then plays it forward in four scenarios.

Countries become self-contained, not self-sufficient. Each one builds the layers it can control (payment rails, data localisation, local models, defence drones, power) and depends on others for the layers it cannot yet make: leading-edge chips, AI memory, rare-earth processing, frontier models and, for many, fuel.

The rest of this brief uses that frame. Section 1 tests it pipeline by pipeline, section 2 shows the resulting stacks, section 3 runs it forward to 2030.

Seven pipelines, September 2026

Each of these is being rebuilt at the same time, which is why the period feels so unstable. For each: what Washington is doing, what Beijing is doing, and what it signals.

Money rails

Betting on privately issued dollar stablecoins instead of a central bank digital currency. The GENIUS Act (July 2025) requires one-for-one backing in dollars and short-term Treasuries. Treasury published its proposed rule on 17 August 2026; most implementing rules are still unfinished.

Stablecoin supply has passed $300bn, and issuers are now significant buyers of US Treasury bills. Every digital dollar issued abroad extends dollar reach and helps fund US debt.

Two layers. CIPS, the renminbi clearing network, hit a record daily average of about RMB 920bn (~$134bn) in March 2026, with usage rising after the Iran war. mBridge, a multi-central-bank digital currency platform, is moving to commercial launch through a Hong Kong entity at roughly half the fees of conventional rails; the digital yuan makes up about 95% of its volume.

A setback: Saudi Arabia has just left mBridge. Around 30% of China's own trade now settles in renminbi, up from 10% in 2017.

At the New Delhi BRICS summit (12–13 Sept) members approved India's UPI-style "BRICS Pay" proposal as a task force to study linking national payment systems. India's sherpa ruled out a common currency. UPI itself processed 24.5bn transactions in August.

Signal: SWIFT is not being replaced so much as flanked. The dollar is going digital through private tokens; the renminbi is growing in commodity trade and in corridors exposed to sanctions. Cost and sanctions risk decide which rail a trader picks.

Energy

Iran declared the Strait of Hormuz closed on 4 March 2026; the IEA called it the largest supply disruption in oil market history. Qatar declared force majeure on LNG. A June US–Iran memorandum collapsed in July, and in September drone attacks shut Saudi Arabia's East-West pipeline, a route that carries up to 4% of world oil.

Already the largest producer of oil and gas and the largest LNG exporter. In 2025 it supplied 31 of the 33 bcm of global LNG export growth. In June 2026 Asia overtook Europe as the biggest buyer of US LNG for the first time since 2024. Energy purchases now appear inside trade and investment deals.

Imported 86% of the oil it consumed in 2025 (China 73%, Europe 75%). China's answer, from its longer-running strategy, is electrification plus domestic coal, renewables and pipeline gas. India is further behind on that path.

Signal: Energy has become a security layer. Importers diversify toward US cargoes and electrify; exporters without an alternative route to the sea lose leverage.

Chips, compute and data centres

Nvidia holds an estimated 80–85% of data-centre AI accelerators. Policy toward China has swung repeatedly: H200 sales were allowed with a 25% levy and volume caps, then Beijing blocked imports. Analysts put the US lead in AI compute produced in 2026 at somewhere between 21 and 49 times China's.

Chips are now diplomacy. The UAE can import up to 500,000 advanced Nvidia chips a year under security conditions; the first 200 MW of Stargate UAE is due this year. Saudi Arabia's HUMAIN has committed about $100bn. Both run entirely on US-licensed silicon.

Domestic chips took about 41% of China's AI-chip market in 2025, roughly half of it Huawei. DeepSeek's V4 was reportedly tuned for Huawei's Ascend processors. A Huawei-led coalition is aiming for 70% semiconductor self-sufficiency by 2028.

The hard limits are advanced manufacturing nodes and high-bandwidth memory. HBM from SK Hynix and Micron is sold out through 2026 even for Western buyers.

Data-centre capacity reached about 1.64 GW in June 2026 and is projected at 6 GW by 2029, needing about $110bn. Global hyperscalers have committed over $50bn, helped by a tax holiday to 2047 for foreign cloud providers. The IndiaAI Mission has onboarded about 38,000 GPUs on an outlay near $1.25bn, against roughly $80bn from Microsoft, AWS and Google. The buildings are local; the chips, memory and most of the platforms are imported.

Private AI investment, 2025

India's data-centre capacity

Signal: This is where your thesis is strongest. The dependency sits in silicon, memory and cloud platforms, and building data centres locally does not remove it.

Rare earths and critical minerals

April 2025 controls on seven medium and heavy rare earths remain in force. The sweeping October 2025 rules, which required a licence for any foreign product containing 0.1% Chinese-origin rare earths, are only suspended, until 10 November 2026. In June 2026 Beijing added US rare-earth producers to its export-control list. It also used rare-earth restrictions against Japan this year over Taiwan.

Building an ex-China chain, anchored on an October 2025 critical-minerals framework with Australia. Progress is slow: China shipped just 17 tons of yttrium to the US in the eight months after controls, versus 333 tons in the eight months before, and aerospace makers report rationing.

Signal: The nearest hard decision point in this entire brief is six weeks away.

AI, jobs and the wealth gap

From 2025 to 2030: 170m jobs created, 92m displaced, net +78m. About 39% of core skills change. Of every 100 workers, 59 need retraining and 11 are unlikely to get it. The largest absolute growth is in frontline work: farming, delivery, construction, care and teaching.

Geoeconomic confrontation (sanctions, tariffs, weaponised supply chains) is the top risk for both 2026 and 2028. Economic downturn and inflation each climbed eight places; asset-bubble risk climbed seven. Adverse outcomes of AI rose faster than any other risk. Half of respondents expect turbulent or stormy years ahead.

No aggregate jump in unemployment yet among AI-exposed workers. The damage is concentrated at entry level: workers aged 22–25 in the most exposed occupations saw about a 13% relative decline in employment. AI was cited in 116,175 of 529,914 announced US job cuts from January to August 2026.

Stanford's AI Index 2026 adds that employment of software developers aged 22–25 has fallen nearly 20% since 2024, even as 88% of organisations now use AI in at least one function.

Jobs created and displaced, 2025–2030

Signal: The first rung of the career ladder is being removed before whole jobs are. The wealth gap widens through two channels: returns to whoever owns compute and models, and a rising premium for people with AI skills.

Robotics

Officials expect humanoid robot output to pass 100,000 units in 2026. Unitree and AgiBot are projected to take nearly 80% of shipments. Prices fell by more than half year on year. China also installed 54% of the world's new industrial robots, up from 51.1% in 2023.

Tesla's Optimus line at Fremont started this summer with output Musk called "extremely slow"; true mass production waits for a Texas plant in 2027. US firms lead on AI and capital, China on manufacturing volume.

Signal: Robots follow the same split as chips, with the roles reversed: China leads on hardware volume, the US on the intelligence inside.

Warfare

By early 2026 Ukrainian firms were making up to 1,000 interceptor drones a day. A Shahed drone costs about $20–50k; a Patriot PAC-3 interceptor about $4.1m. When Iran launched hundreds of Shaheds at Gulf states in March, advanced air-defence stockpiles drained fast. Newer Shahed variants carry seekers that recognise targets on their own in the final kilometres.

Iranian missiles struck the Gulf in February 2026, exposing a gap in the sovereign-AI debate: a data centre is a physical asset that now needs air defence. The same applies to pipelines, LNG terminals and subsea cables.

Signal: Drone manufacturing is becoming a sovereign capability in its own right, and AI moves into targeting and interception first, ahead of fully autonomous swarms.

Two stacks, and what India owns or rents

Put the pipelines together and two technology-and-money stacks appear. Most countries will plug into both and own only some layers themselves.

Four scenarios

The likelihoods are my rough judgement from the evidence above, meant to show relative weight. They are not forecasts, and pieces can combine (a hard split can also trigger a financial reckoning).

A · Two stacks, many plugs

B · Hard split

C · Acceleration

D · Reckoning

How governments may respond: a ladder with triggers

Google DeepMind economists Julian Jacobs and Alex Imas published an essay in July 2026 (updated September) setting out three economic futures for advanced AI and a policy response for each. A response switches on only when specific data crosses a line. The essay states it is the authors' view and not Google's official position. Their three futures map onto the scenarios above.

They rate eleven options, from wage insurance and a jobs guarantee to universal basic income and a sovereign AI dividend modelled on Alaska's oil fund. On retraining, they find public programmes have a disappointing record and apprenticeships and employer-led models do best. Like the labour data in section 01, they find no definitive evidence yet of economy-wide job or wage losses.

A way to read the transition from data instead of headlines. Watch labour's share of GDP, median wages and how long people stay unemployed.

A gap that sits exactly on your thesis. The essay is written for a country that owns its AI companies, and it does not discuss cross-border issues. A sovereign AI dividend or a public capital fund is paid for by taxing or owning AI profits. In a rented-compute country, much of that profit is booked by foreign platforms, so the tax base sits abroad. That is my reading, not the authors'. For India it turns digital-services taxation, data localisation and public stakes in domestic AI infrastructure into welfare questions as well as industrial ones.

Dates that tell you which scenario you are in