The AI bill is going up, and Asia sits on both sides of it

Most AI budgets I see in the region rest on one quiet assumption: whatever this costs today, it will cost less next year. Token prices fell for three years straight, and every finance team learned to treat the line as deflationary.

That assumption had a bad week. Between 29 September and 5 October, the price of renting compute went up, the price of the memory inside that compute went up, a regional macro watchdog said Asia would take the worst of any AI correction, and Korea's banks discovered what it costs to defend against attackers using the same tools.

None of these is a crisis on its own. Together they say the cost curve has turned, at least for now. Asia is unusual because it sits on both sides of the bill. It makes the scarce parts, and it buys the finished service.

The rent on compute just rose by a fifth

On 1 October Nebius moved its on-demand rates for NVIDIA GPUs. An H100 went from $3.85 to $4.50 per GPU-hour, an H200 from $4.50 to $5.40, a B200 from $7.15 to $8.50 and a B300 from $7.85 to $9.50, according to its own price list. That is a rise of 17 to 21 per cent, and the second rate-card increase in roughly three months. The same page offers up to 35 per cent off for reserving large clusters for several months, which is the other half of the message: commit now or pay more later.

The retail end moved the same way. On 29 September OpenAI reopened its $200 Pro plan with roughly half the old API-dollar allowance, with existing subscribers grandfathered only until 29 October. Same sticker price, less inside the box.

So what: If your AI cost model assumes a falling price per task, rebuild it with a flat or rising one and see which projects still clear. The ones that only worked on a falling curve were never really approved.

Korea is selling the scarce part at five times the contract price

The reason compute is getting dearer is partly memory, and memory is largely Korean. Spot prices for a 36 gigabyte HBM3E stack reached about $2,100, four to five times contract levels, according to market research firm MegaGrid Supply as reported by the Seoul Economic Daily on 31 August. The same report cites Korea International Trade Association data showing DRAM export volume fell 13.2 per cent between May and July while export value rose 18.5 per cent, as the average export price jumped 36.6 per cent. Analyst consensus has Samsung's third-quarter operating profit at 116.38 trillion won and SK hynix's at 79.16 trillion won.

That is a seller's market, and the seller is in this region. Korean suppliers are capturing margin that a year ago sat further down the chain.

So what: For buyers, hardware-heavy plans such as on-premise inference clusters or sovereign compute projects should carry a memory price contingency, not a fixed quote. For suppliers and their lenders, record margins at the top of a cycle are exactly what the next section is about.

AMRO says the region is long the boom

On 5 October the ASEAN+3 Macroeconomic Research Office released its 2026 Financial Stability Report. The region accounted for about two-thirds of the growth in global AI-related trade, and AMRO's AI slowdown scenario would cut ASEAN+3 growth in 2027 to 2.6 per cent from a baseline of 4.1 per cent, the largest of the shocks it modelled (The Nation, 5 October 2026). It named Korea's AI-concentrated equity market and the close tracking of Japanese and Hong Kong stocks to American tech, and flagged hyperscalers' growing use of debt to fund data centres (Free Malaysia Today, via Bloomberg). Its lead economist, Runchana Pongsaparn, put it carefully: "There are some warning signals, but we would say that they are not grave yet."

The capital picture underneath is narrow too. Tech Collective, drawing on e27 data, counts US$2.81 billion of Southeast Asian startup funding in the first quarter across 98 deals, with about 70 per cent of it from one $2 billion data centre round for DayOne.

So what: A region whose growth, equity markets and venture money all lean on the same build-out should not also let its operating plans lean on it. If your AI programme only makes sense while compute is scarce and valuations are high, it is a bet on the cycle, not a capability.

Korea's banks found the third bill

The cost nobody budgets for is defence. Starting with Shinhan Bank, where 25,727 customers were affected, a run of suspected AI-assisted intrusions spread through Korean finance. Yegaram Savings Bank lost data on about 40,000 customers, the largest single breach, and KB Kookmin, Hana and BNK Busan reported smaller incidents. Detection took Shinhan 15 hours 26 minutes, Hana 41 hours 44 minutes and KB Kookmin 67 hours 41 minutes. Investigators linked the attacks to a Chinese-language tool called Artex AI. On 4 October the Financial Services Commission called an emergency meeting and the Financial Supervisory Service alerted around 500 financial firms, with checks due by 9 October (Korea JoongAng Daily, 4 October 2026).

So what: When attack costs fall, defence costs rise. Any business case that counts AI savings on the operations side and ignores the extra security spend it forces is incomplete, and regulators in the region are now asking for evidence, not policies.

The through line

Cheap AI was a phase, not a law. The inputs are getting dearer because Asia's own suppliers have pricing power, the services built on them are being repriced quietly, and the region's macro exposure means a turn in the cycle would land here first. None of this is a reason to stop. It is a reason to budget like a captain who checks the weather rather than one who assumes the tide only comes in. Model your AI spend at flat or rising unit costs, lock in terms while you have leverage, keep an open-weight fallback for workloads you could self-host, and put the security bill on the same page as the savings.