SemiAnalysis has sharply reduced its long-range forecasts for global AI accelerator spending, chip shipments, high-bandwidth memory demand and electricity use, concentrating the largest changes in 2028 through 2030. Its updated model projects $923 billion in AI accelerator capital expenditure in 2030, down $608 billion from its previous $1.531 trillion estimate.
The Nov. 13 update points to a more constrained buildout of AI computing infrastructure after 2027, driven by slower growth in annual recurring revenue at frontier AI labs, weaker financing appetite for data center projects without hyperscaler commitments, and expected construction delays in the United States.
SemiAnalysis kept its 2026 accelerator-capex estimate broadly unchanged at roughly $392 billion. The reductions grow substantially in later years: the firm lowered its 2027 estimate to $627 billion from $770 billion, a 19% decrease; cut 2028 to $758 billion from $1.054 trillion, down 28%; and reduced 2029 to $796 billion from $1.352 trillion, a 41% decline.
Those revisions do not imply that AI infrastructure spending is expected to fall in absolute terms. The updated model still projects annual expenditure rising from 2026 to 2030. The change is that SemiAnalysis now expects a far slower expansion than previously forecast, particularly after the first wave of large-scale data center construction.
Nvidia takes most of the wafer-shipment reduction
Nvidia accounts for 66% of the total reduction in projected AI accelerator logic-silicon wafer shipments in the new model, according to SemiAnalysis. The firm cut its cumulative 2026-2030 forecast for the overall AI accelerator wafer area shipped by 17%.
Annual reductions in the wafer-area forecast reached 13% for 2027, 16% for 2028, 24% for 2029 and 15% for 2030. Wafer area provides a way to measure the physical manufacturing demand behind accelerator shipments, capturing the volume of advanced chips expected to be fabricated rather than simply the revenue attached to those chips.
For Nvidia, SemiAnalysis reduced projected compute revenue shipments by 22% in 2027, 39% in 2028, 56% in 2029 and 66% in 2030. On a total-system-power basis, which includes the electricity requirements of deployed AI systems, its shipment projections fell by 14%, 35%, 50% and 59% over the same years.
The firm also cut assumed average selling prices for Nvidia's future Feynman products. SemiAnalysis now models Feynman at approximately $78,000, down from about $110,000; Feynman Ultra at about $98,000 rather than $175,000; and Feynman Next at roughly $118,000 instead of $191,000.
Lower projected pricing changes the revenue outlook even if unit shipments remain relatively resilient. It also suggests SemiAnalysis expects hardware buyers to gain greater purchasing leverage, while competing products and custom silicon reduce the ability of a single supplier to sustain exceptionally high system prices over several hardware cycles.
Memory and power requirements also decline
The forecast revision reaches beyond accelerators themselves. SemiAnalysis lowered its estimate for global high-bandwidth memory, or HBM, shipments in 2028 to 6.4 exabytes from 9.0 exabytes, a 30% reduction. HBM is the specialized memory packaged close to AI processors to move large datasets at high speed.
Its 2029 HBM forecast was reduced to 7.9 exabytes from 10.3 exabytes, a 23% cut. The 2030 estimate fell to 9.6 exabytes from 10.9 exabytes, down 12%.
The smaller reductions in 2030 memory volumes than in AI spending may reflect expectations that memory capacity per system will keep increasing, even under a more moderate accelerator deployment curve. AI workloads remain memory-intensive, and future systems may rely on larger memory configurations to support models with longer context windows, more users or more complex inference tasks.
SemiAnalysis also cut its estimate for AI accelerator electricity demand. It now projects 80.7 gigawatts of power demand in 2030, down 34% from its prior 122.8-gigawatt forecast. The 2029 projection fell to 72.5 gigawatts from 108.7 gigawatts, a 33% reduction.
A gigawatt is roughly the output of a large power plant, making the revisions material for developers planning data center campuses, power infrastructure and grid connections. Less projected equipment deployment would reduce the near-term urgency of some power projects, though it would not remove the bottlenecks surrounding transmission, generation and permitting in major U.S. data center markets.
Data center delays reshape the deployment schedule
SemiAnalysis incorporated an assumption that 5.8 gigawatts of U.S. data center supply originally expected between the second half of 2027 and the first half of 2028 will be delayed because of U.S. midterm election outcomes.
The forecast also cites slower third-quarter growth in new annual recurring revenue at frontier AI labs and reduced appetite from debt and equity providers for infrastructure projects that do not have purchase commitments from major cloud companies. These factors place greater weight on contracted demand rather than speculative capacity construction.
That distinction affects hardware suppliers differently. SemiAnalysis reduced Broadcom's cumulative 2026-2030 ASIC revenue forecast by 11%, attributing much of the decline to lower HBM-related cost and pricing assumptions rather than deep shipment cuts. ASICs are application-specific integrated circuits, custom chips designed for a particular customer or workload.
The firm's forecasts for AMD were cut more deeply after 2027. On a total-power basis, SemiAnalysis reduced AMD's 2028, 2029 and 2030 outlook by 23%, 55% and 57%, respectively. Spending forecasts were lowered by 29%, 35% and 32% for those years. The 2027 total-power forecast was nearly unchanged, while the spending estimate fell 11%.
SemiAnalysis also shifted expected deliveries of about 25,000 MI400 chips and 30,000 Meta-custom MI450-MC chips from the fourth quarter of 2027 to the fourth quarter of 2026. The timing adjustment brings a portion of anticipated AMD deployment forward rather than removing it entirely.
Custom silicon outlook becomes more selective
For Google, SemiAnalysis reduced lifetime Humufish shipments by 470,000 units across 2028 and 2029 and cut lifetime projections for the Triggerfish program aimed at Anthropic by 42%. By contrast, the firm left its 2027 and 2028 Trainium forecasts unchanged, citing Amazon's data center project inventory and the lower manufacturing cost of Trainium.
The mixed treatment of custom silicon shows that the revised outlook does not assume an even pullback across all chip programs. Hardware tied to existing cloud capacity and lower-cost designs may retain stronger demand visibility than projects dependent on new financing or uncontracted data center construction.
The update also included Groq LPU and standalone CPU products within its revised scope, extending the reassessment beyond the largest GPU suppliers.
Crypto narratives face a weaker hardware-growth backdrop
The forecast changes provide a less supportive backdrop for crypto assets marketed around AI computing, decentralized GPU networks or data-center capacity. Those tokens often trade on expectations that demand for AI compute will expand rapidly and create a need for alternative sources of processing power.
SemiAnalysis's projections do not directly measure demand for decentralized networks, and they do not establish a price outlook for related digital assets. They do indicate that the market for physical AI infrastructure may develop more gradually than many high-growth narratives assumed.
For token traders, the most useful distinction is between projects with identifiable computing usage and those whose valuation largely depends on broad enthusiasm for AI hardware scarcity. A slower pace of accelerator purchases, HBM deployment and data center power buildouts could make it harder for the latter group to sustain claims that centralized supply shortages will automatically drive demand toward token-based alternatives.
For deeper context on AI’s role in markets and trading, explore our guide to how AI complements blockchain today.
Disclaimer: The content on this page is provided for general informational purposes only and does not represent the views or financial advice of Toobit. We make no guarantees regarding the accuracy or completeness of this information and shall not be held liable for any errors, omissions, or outcomes resulting from its use. Investing in digital assets involves risk; users should independently evaluate their financial situation and the risks involved. For further details, please consult our Terms of Service and Risk Disclosure.
