Bangladesh's Power Glut and the AI Hiring Reversal: Two Faces of a Capacity Mismatch
On 9 June 2026, Nikkei Asia reported that Bangladesh's decade-long build-out of power generation capacity has left the country with more electricity than its grid can absorb or its treasury can subsidise. Plant after plant sits idle or runs well below nameplate, while the state still owes those same plants capacity payments and the government still owes the fuel suppliers and the lenders. The result, as Nikkei documented, is a paradox: the unit cost of electricity has risen rather than fallen, and the fiscal subsidy bill has ballooned, even as the average consumer waits hours for a stable connection.
On the same day, a separate datapoint landed from a survey circulated by Unusual Whales: among organisations that laid off staff on the assumption that artificial intelligence would absorb the work, roughly 38 per cent cited the technology's higher-than-expected oversight and quality-control requirements as a primary reason for rehiring. The two stories sit in different geographies and different sectors, but the diagnosis is identical. Both are cases of capacity being financed on the assumption of a future demand profile that did not materialise — or that arrived with costs the planning models did not price in.
Bangladesh: a power system priced for a tiger that slowed
The Nikkei account sketches a familiar sequence in developing-country industrial policy. Over the past fifteen years, Dhaka — supported by multilateral lenders, Chinese contractors, Indian equipment vendors and a constellation of independent power producers — added generation capacity at a pace that assumed double-digit GDP growth, continuous export expansion from the garment sector, and a residential demand curve that would rise with urbanisation. The plants were built. The growth arrived, but more slowly and more unevenly than the power purchase agreements (PPAs) had priced in.
When a plant runs below the utilisation threshold written into its PPA, the off-taker — in Bangladesh, the Bangladesh Power Development Board — is still obliged to pay. The bill for these so-called "capacity payments" is one of the two structural pressures Nikkei identifies. The other is the fuel cost of running plants that are technically available but economically marginal: gas and oil imports that the treasury must finance through a subsidy account that has, in recent years, been restructured more often than it has been cleared.
The political economy of the build-out is worth naming. The plants were not built by accident. They were built because domestic constituencies — the local engineering, procurement and construction firms, the foreign suppliers with their own export-credit agencies behind them, the politicians who cut the ribbons — all had reasons to add capacity faster than the grid could absorb. The multilateral lenders, for their part, financed deals whose risk profile looked tractable in a high-growth scenario and brittle in a low-growth one. The result is a system in which the marginal kilowatt costs more than the average kilowatt, the state pays for power it does not use, and consumers still face routine outages. The Nikkei reporting makes the point carefully: this is not a failure of ambition but a failure of calibration.
The AI hiring reversal: when the tool's own overhead becomes the constraint
The Unusual Whales figure — 38 per cent of those who cut staff because of AI citing higher-than-expected oversight and quality control as a primary rehiring driver — describes a structurally similar error on a much shorter time horizon. The build-out, in this case, was a workforce plan: the assumption that large language models, retrieval systems and agentic workflows could absorb the work of a category of employee at a quality level that did not require proportionally more human supervision.
The evidence so far suggests that for a meaningful slice of tasks, that assumption was wrong. AI systems do not, in general, eliminate the human-in-the-loop; they shift the loop's centre of gravity. Reviewers spend less time producing the first draft and more time checking for hallucination, drift and edge cases. Compliance, legal and customer-facing functions have discovered that a chatbot's answer is only as cheap as the audit trail that defends it. The 38 per cent figure, if the survey methodology is sound, is a measure of the gap between vendor roadmaps and operating reality.
The mechanism is the same one Bangladesh's power sector illustrates. A capacity is contracted for on the basis of a unit-economics projection. The projection turns out to have omitted a category of cost — in Bangladesh's case, the cost of idle capacity and stranded fuel; in the AI case, the cost of supervising, auditing and quality-controlling the model. The buyer then has two options: pay the hidden cost and keep the capacity, or rebuild the workforce and admit that the projection was wrong.
A shared structural frame: capacity contracted for a demand that did not arrive
The temptation in each story is to treat it as a local failure — Bangladeshi planning, in one case; a particular firm's AI strategy, in the other. The structural reading is more useful. Both are instances of a system that priced a future and got the price wrong, and both are now stuck with a sunk-cost overhang that distorts the unit economics of the present.
The pattern is not new. Rail over-builds in the nineteenth century, fibre over-builds in the late 1990s, Chinese solar over-builds in the 2010s, and the global LNG ordering spree of 2022-23 all share the shape: a forecast of demand that looked reasonable when the contracts were signed, a slower or differently-shaped arrival of that demand, and a fiscal or corporate balance sheet left carrying the difference. What is distinctive about the current moment is the simultaneity. Energy systems, digital systems, and the labour markets that wrap around them are all recalibrating at once, and the political cover for admitting miscalculation is thin in each.
Stakes: who pays for the forecast that missed
In Bangladesh, the answer is the consumer and the treasury. The consumer pays through a tariff structure that is politically difficult to fully pass through, and through the reliability cost of a grid that is technically capacious but operationally fragile. The treasury pays through the subsidy account, which has been a recurrent source of fiscal stress and which the International Monetary Fund's programme engagement has repeatedly tried to restructure. The plant owners, in many cases protected by sovereign-guarantee PPAs, are paid first.
In the AI case, the answer is the firm that over-cut and the worker who was hired back at a different price. The 38 per cent who rehired discovered that the all-in cost of AI-augmented work was not the marginal cost of compute; it was compute plus the human supervision that the model still required. The wage bill, in many cases, is being rebased upward, not downward. The broader labour market consequence is harder to read: the firms that learnt the lesson quickly may be reabsorbing the same people at the same or higher wages; the firms that over-cut and under-rebuilt may simply be operating below their previous quality level.
The forward view, in both cases, is for a slower, more cautious expansion of capacity. New Bangladeshi power plants will, presumably, be procured with stricter dispatch discipline and less generous take-or-pay terms; new AI workforce plans will, presumably, be drafted with more conservative assumptions about what the technology can absorb unaided. The more interesting question is what happens to the capacity that is already locked in. The idle plant and the rehired worker are both monuments to a forecast that was, briefly, very widely shared.
What the sources do not tell us
The Nikkei summary circulated on Telegram does not, in the version available, give the full numerical range of Bangladesh's idle capacity or the specific composition of the subsidy backlog; it frames the diagnosis rather than the ledger. The Unusual Whales figure of 38 per cent is a single survey datapoint without methodology disclosed in the circulating excerpt. In both cases, the underlying numbers would benefit from the primary documents — the Bangladesh Power Development Board's annual report, in one case, and the survey instrument and sample frame in the other. The structural reading offered here does not depend on those numbers being exactly right, but the precision of any forward claim does.
Desk note: Monexus has run the two stories together because the underlying mechanism — capacity contracted for a demand that did not arrive, with the shortfall paid in a different currency than the planners expected — is identical. The Bangladesh piece leans on Nikkei Asia's reporting; the AI figure is treated as a survey signal pending primary corroboration. We have not speculated beyond the source material on either the unit-cost figures or the rehiring methodology.
Wire provenance
This editorial synthesis draws on the following public wire/social posts:
- https://t.me/s/NikkeiAsia
- https://t.me/s/nikkeiasia