Eight hundred and sixty billion is a round number in won

The figure moving this morning is that the market for AI chips used in data centres will reach $860bn by 2030. A good number: large, round, four years out and detached from any company you could check it against. So begin with the author, because a forecast is a claim and claims have owners.

This one belongs to the Overseas Economic Research Institute at the Export-Import Bank of Korea, reported on Sunday by Kim Hye-ran in the Seoul Economic Daily. The institute projects the data-centre AI chip market growing from $124bn in 2024 to $860bn in 2030, an average of 38 per cent a year, with demand shifting from training to inference and so towards neural processing units, which are cheaper and draw less power. It puts Nvidia at 78.2 per cent of the market last year, Google at 4.7, AMD at 4.1, Intel at 3.7 and Huawei at 2.6.

The arithmetic is honest, which is the first thing to say: $124bn to $860bn over six years requires 38.1 per cent a year, so the headline rate is the rate.

The currency is where the precision starts to dissolve. The institute gives both figures in won as well: 170 trillion won in 2024, 1,200 trillion in 2030. Those imply 1,371 won to the dollar in the base year and 1,395 in the terminal year — two different exchange rates in a single sentence, neither stated as an assumption. What the institute has actually forecast is 1,200 trillion won, a round number in its own currency. The $860bn is a conversion of a round number, presented to three significant figures, and a forecast four years out does not get more accurate by crossing a currency.

Then the definition, which is absent. "AI chips used in data centers" is defined nowhere in the coverage. Does it include high-bandwidth memory? Networking silicon? Whole systems, or the accelerator alone? The share table answers part of that in a way that raises more: Google is credited with 4.7 per cent, and Google sells tensor processing units to nobody. To give it a share you must price an internal transfer — a defensible methodological choice, and an undisclosed one.

Now the test that matters, which is against a filing rather than against another forecast. Nvidia's quarterly report for the three months to 26 July 2026 puts Data Center revenue at $89.023bn, up 117 per cent on the year, and $164.269bn for the six months. One company, in half a year, booked more revenue than the institute's entire market in its base year. Annualise the July quarter and Nvidia alone is running at $356bn. Apply the institute's own 78.2 per cent share and the 2026 market is somewhere near $455bn — more than half the 2030 figure, reached four years early.

Run that forward and the forecast inverts. From roughly $455bn in 2026 to $860bn in 2030 is growth of about 17 per cent a year, not 38. This is not an aggressive projection. It is a deceleration forecast with a two-year-old base, wearing a historical growth rate as a headline: the 38 per cent describes what already happened, and most of the distance from $124bn to $860bn was covered before the forecast was published.

One caveat against this desk's own test, stated because it cuts the other way. Nvidia's Data Center segment includes networking and complete systems, so it is not a like-for-like measure of chips and the implied 2026 market is an upper bound. That is exactly the difficulty: with no published definition, there is no honest way to reconcile the forecast with any company's accounts. A number that cannot be checked against a filing is neither conservative nor aggressive. It is untestable.

Which brings the question back to why it exists. The newspaper's own subheading says it: "Export-Import Bank Urges Support for Early Demand, Overseas Expansion". The institute's conclusion is that Korean entrants — FuriosaAI, whose RNGD chip went into mass production early this year, and Rebellions, whose REBEL 100 is due in the second half — have the design expertise but trail the leaders "by more than three years in capital, commercialization experience and ecosystem". Its recommendations are tax breaks or subsidies for buyers of domestically made AI chips, and diplomatic support for export sales, with the Middle East singled out because governments there want alternatives to expensive Nvidia hardware.

That is a coherent industrial policy argument, and it is the purpose the forecast serves: the institute is the research arm of a state-owned policy bank whose business is financing Korean exports, and the bigger the 2030 market, the larger the prize from getting two domestic startups into it. None of that makes the number false. It makes it an advocacy number, to be read with the proposal attached — which the wire copy detaches. Two caveats on sourcing, in the same spirit: the institute's own publication could not be reached from this desk, so the method is known only through a newspaper summary, and that newspaper labels its English edition as AI-translated from Korean.

Set it beside the two forecasts this desk has tested in the past week. The Brookings paper on financing the build-out publishes its comparison table, its gigawatt assumptions, the share of the pipeline it assumes never gets built, and the single figure that would falsify it: about $5.50 of revenue per installed GPU-hour by 2032. Micron's take-or-pay book is demand already contracted, cash paid in advance, terms in a filing. Against either, $860bn is a destination with no route published — and the question worth asking of any projection is the one this one cannot answer: what would have to happen for it to be wrong?