An analysis says the artificial intelligence (AI) bubble could burst when supply and capital investment outpace demand that can be monetised, rather than due to technology failure.

On Aug. 8, IT outlet SiliconANGLE reported that bottlenecks in the AI supply chain remain in memory, advanced packaging, networks, power and site readiness, delaying the timing of any bubble collapse.

The key is not a lack of demand but a shift in bottlenecks. Even if firms secure graphics processing units (GPUs), deployments are difficult if high-bandwidth memory (HBM) is scarce. Even with HBM, it must go through advanced packaging to become a completed accelerator. Networks, power, sites and funding must also be in place before capacity can generate profits. The point is not that AI stops working and the bubble bursts, but that if utilisation and cash flow fail to keep pace before scarcity eases, the capital cycle can turn down.

Market indicators also show this structure. In 2024, David Floyer (데이비드 플로이어) projected that the expanded silicon ecosystem would approach $1 trillion by 2028, but the market is moving faster. Global semiconductor sales in 2025 neared $800 billion, and the World Semiconductor Trade Statistics (WSTS) forecast 2026 sales at about $1.51 trillion. Nvidia posted $75.2 billion in data centre revenue in its latest quarter, Broadcom recorded $10.8 billion in AI semiconductor revenue, and AMD reported $5.8 billion in data centre revenue, supporting the view that AI demand is unusually strong.

More than half of the projected market in 2026 is accounted for by memory. With structural demand compounded by price increases driven by scarcity, the revenue curve is rising more steeply than physical shipments and actual deployment speed. Micron's recent results are a clear example. DRAM bit shipments rose only by a low single-digit percentage from the previous quarter, but average selling prices (ASP) per bit rose by the low 60 percent range. If prices fall but bit demand and productive use keep rising, that can be seen as normalisation. If falling prices coincide with slowing shipments, it is a warning signal that supply is outpacing effective demand. The analysis says additional HBM supply must also pass through validation, vertical stacking and advanced packaging, so investors should watch bit shipments and ASP together with package lead times and yields.

A risk factor is that AI factories move to two different clocks. The short-term IT clock, made up of GPUs, custom accelerators, HBM, networking and servers, allows quarterly orders and deliveries and is refreshed on a three to six-year cycle. The long-term site clock, covering sites, substations, grid connections, cooling and water, is used for decades and can take about 10 years for permits and construction alone. Even if capital is invested and hardware is allocated and suppliers report strong order results, installations must pass through installation, acceptance, power supply and productive use before they translate into sustainable revenue and cash flow. The analysis says bubble risk builds in this time lag.

Oracle, OpenAI and the Stargate project were presented as representative examples showing the gap between commitments and actual revenue generation. OpenAI was reported to have committed to buy $300 billion worth of Oracle computing capacity over five years as part of Stargate construction starting in 2027. Oracle increased capital expenditure in fiscal 2026 to $55.7 billion, up 162 percent from a year earlier, and said it could spend up to $95 billion in gross terms and about $70 billion in net terms in fiscal 2027. Oracle's remaining performance obligations (RPO) reached $638 billion, but the share expected to convert into revenue within 12 months was only about 12 percent.

OpenAI's forward compute commitments through 2030 are estimated at $600 billion to $665 billion, which is tens of times its current annualised revenue, and cash burn is continuing. Oracle's free cash flow swung to a $24 billion deficit in fiscal 2026 from a surplus of about $26 billion in fiscal 2025, and its credit rating slid to the borderline of investment grade as debt increased. The analysis says commitments show intent but do not prove actual deployment, productive use or capital recovery.

Intel and China were cited as variables. Intel, under Chief Executive Lip-Bu Tan (립부 탄), has reduced much of the bankruptcy concern raised under former CEO Pat Gelsinger (팻 겔싱어), helped by a restructuring that cut more than 20,000 jobs, U.S. government equity participation and strategic investments from Nvidia and others. Still, it has no flagship customer yet that has secured external foundry volumes, raising the possibility that capacity expansion reliant on policy funding could be prolonged without commercial validation. China, by contrast, was cited as a variable that could bring forward the timing of global price normalisation by expanding supply in mature nodes, NAND and commodity DRAM and by increasing use of its own AI models and platforms despite broader gaps in advanced technologies.

The analysis laid out three future scenarios. One is a soft landing in which demand absorbs supply without strain and memory prices normalise gradually. Another is a delayed clearing in which bottlenecks are not resolved and keep shifting, extending the boom beyond 2027 while widening the gap between committed capital and cash-generating capacity. The third is a bubble collapse in which productive use fails to follow expanding supply and falling prices, eventually blocking capital raising.

Early warning indicators cited include ASP relative to HBM bit growth, advanced packaging capacity, lead times and yields, GPU rental prices and cluster utilisation rates, actual powered capacity versus announced capacity, order backlogs and customer prepayments, and free cash flow. If these indicators weaken at the same time, it can be interpreted as a signal that scarcity is turning into surplus.

The analysis pointed to 2028-2029 as a high-risk window. It cited 2029 in particular as a time when bottlenecks ease significantly, China's production capacity expands sharply and Intel is likely to become a practical alternative to Taiwan's TSMC, making it a test year for whether supply outstrips profitable demand. It projected that if delays in power infrastructure persist, the risk window could be pushed into the 2030s.