AI Isn’t Slowing Down. The Money Is Just Moving.

Compute demand is not disappearing: where AI capital is going and how demand is shifting

On September 14, 2026, a selloff hit U.S. AI stocks.

Semiconductor shares fell across the board. ARM dropped 9.74%. ASML fell 7.25%. Intel declined 5.59%. Broadcom slid 4.77%. AMD fell 4.4%. TSMC dropped 3.41%. Memory stocks also fell. SK Hynix lost more than 7%. SanDisk fell nearly 5%.

Cybersecurity stocks moved the other way. Palo Alto Networks and CrowdStrike each rose more than 13%. Okta, Zscaler, Qualys, SentinelOne, Netskope, and other security companies posted double-digit gains.

The trigger was an article by Anthropic CEO Dario Amodei. He called for slowing the pace of frontier AI model development. OpenAI CEO Sam Altman and SpaceX CEO Elon Musk publicly agreed. Amodei did not call for stopping AI work. He said that even with a slower pace, AI progress would still “look very fast.” Altman later wrote that “controlling the pace of development” does not mean “stopping.” It means AI progress may be slower than it could have been, and safety assessments, monitoring, and other measures will add cost.

The market heard the first half and missed the second. SoftBank, OpenAI’s largest outside shareholder, fell nearly 11%.

Ben Reitzes of Melius Research said: “These AI companies may be very good at building models, but they are not good at talking about stocks. They are disrupting the market.”

A harder question sits behind the market reaction. Emmanuel Cau, head of European equity strategy at Barclays, asked what the market needs to answer: Do higher AI safety requirements mean slower and lower capital spending?

Long-term financing will not derail because of one weekend’s news

The first market reaction was panic selling. Institutional investors were calmer.

Thierry Taglione, senior fixed-income investment strategist at AllianceBernstein, said of hyperscaler and data center financing: “These are long-term financing plans. Although (slowing the pace of AI development) is a reasonable concern, we are talking about 10-year and longer financing. This will not derail because of one weekend’s news.” The firm still expects top AI hyperscalers to spend more than $1 trillion next year.

Morgan Stanley raised its capital spending forecasts. A year ago, the market expected the five major cloud providers to spend about $450 billion each in 2026 and 2027. Morgan Stanley now expects about $800 billion in 2026 and about $1.16 trillion in 2027. JPMorgan expects Amazon, Microsoft, Google, Meta, Oracle, CoreWeave, and SpaceX to spend $933 billion in 2026, up from $443 billion in 2025. It expects $1.577 trillion in 2027.

Token consumption is growing fast, and that growth is driving the spending. OpenRouter data show global weekly token use rose from about 6 trillion to 28 trillion since January 2026, a 350% increase. Morgan Stanley said this shows AI demand is not just a concept. Enterprise AI applications, agents, multimodal models, and always-on AI services have pushed it into large-scale deployment.

Compute demand is not disappearing, but it is changing direction

In 2026, a major shift appeared: hyperscaler spending on inference passed spending on training for the first time. The industry focus is moving from “training large models” to “using large models.”

TrendForce data show North America’s five largest cloud providers will increase AI training compute by 56% in 2026. Inference compute will grow 122%, more than twice as fast. Yang Yuanqing, chairman and CEO of Lenovo Group, described the shift: “At present, more than 70%-80% of AI compute is used for training, while 20%-30% is used for inference, but in the future this trend will reverse, and AI compute used for inference will account for more than 70%.”

Inference demand is growing for clear reasons. Agents are replacing chat models as the main interaction form. A single task can call a model dozens or hundreds of times, so token use expands quickly. Lower inference costs have not reduced total demand. They have increased usage. Ben Barringer, global head of technology research at Quilter Cheviot, agreed: even if AI model training and releases slow, inference still faces a compute shortage. “Demand still far exceeds supply, so even if growth slows slightly, company revenue is unlikely to be affected.”

Slower training does not mean less training infrastructure. Yang Yuanqing said: “It does not mean the infrastructure needed for training will decrease; on the contrary, it will continue to grow further.” Morgan Stanley estimates that global data center developers face a power gap of about 55 GW. AI competition is shifting from a “GPU battle” to a “power battle.” Even if some training compute growth slows, power, land, and cooling constraints mean projects already under construction must be finished.

The beneficiaries are changing

Cerebras launched the CS-4 system. The company says CS-4 can output 30 times more tokens per second in single-user scenarios than traditional GPU servers. Cerebras processors use a single complete wafer. Nvidia and AMD use multi-chip assemblies. The Cerebras design shortens data travel distance. It uses SRAM instead of DRAM, so it is faster and does not depend on memory chips that are in short supply. Cerebras CEO Andrew Feldman said: “CS-4 achieves industry-leading speed on frontier large models, fundamentally changing the industry paradigm.”

GPUs and inference chips are not replacing each other in a zero-sum way. Nvidia still holds about 90% of training. Inference is the main area of new growth. Google’s in-house TPU chips are expected to see demand grow nearly 80% in 2026. Amazon’s Trainium series is expected to account for more than 40% of its own AI servers. Nvidia is also shifting focus. Jensen Huang introduced “Token Factory Economics” at GTC 2026. It defines tokens per watt as the main measure of competitiveness. Nvidia spent $20 billion to partner with Groq, which makes LPUs for inference.

The split is a division of labor. Inference chips fit stable model structures and very high call volumes. GPUs remain hard to replace when model structures change quickly. JPMorgan analysts said that with compute supply still tight, token consumption for both proprietary frontier models and open-source models will keep growing.

Safety has become a new cost variable

The AI safety debate is becoming a cost variable that can be measured. Barclays estimates that “pace control” will add more than $44 billion to industry compute costs in 2027. That would raise overall costs by about 18%. In 2028, the added cost could reach $76 billion. OpenAI has disclosed that highly capable models must have real-time monitoring for all reinforcement learning training, evaluation, and inference workloads. Monitoring takes about 20% of the monitored inference compute. About 85% of AI lab compute already goes to post-training and inference. As a result, inference and post-training compute demand will rise by 20%.

Safety compliance is creating new compute demand. Amodei’s three-stage safety framework includes third-party “embedded evaluators,” common industry safety standards, and coordinated international regulation. Each step requires extra compute for monitoring, evaluation, and auditing.

The rise in cybersecurity stocks is another sign. SentinelOne CEO Tomer Weingarten said deep, complete monitoring of computer systems is an important way to tell whether AI is deviating from its original intent. If AI capabilities keep improving, especially as agents gain more autonomous execution power, companies will need more AI behavior monitoring, security protection, and risk control.

The credit market dimension

AI capital spending is spreading from stocks into credit markets. Goldman Sachs research published in August 2026 shows that AI-related debt issuance since the start of 2026 has approached $500 billion. Morgan Stanley said AI financing is expanding and changing. Financing has moved from the dollar market into euros, pounds, Swiss francs, Canadian dollars, and other international credit markets. Issuers are no longer only large tech companies such as Microsoft and Amazon. Data center REITs and neoclouds have also started to raise money in high-yield bond markets.

AllianceBernstein counts more than $330 billion in bond issuance tied to major global hyperscalers and data centers so far this year. Peter Boockvaar, chief investment officer at One Point BFG Wealth Partners, warned: “The U.S. economy, stock market, corporate earnings trajectory, and profit margins are all closely tied to AI capital expenditure. Anything that negatively affects these aspects could trigger a selloff.”

AllianceBernstein also said: “not all hyperscalers are created equal. Managers need to favor companies with strong free cash flow and lower leverage, while underweighting those with tighter finances.” The Bank for International Settlements has warned about rising leverage among tech giants. As investment in data centers, advanced chips, power, and cloud infrastructure grows quickly, tech companies rely more on debt and other financing tools.

Conclusion

The September 14 market swing repriced where value sits in the AI supply chain. It did not mean compute demand will disappear.

Capital spending will not stop because of safety calls. Hyperscaler financing plans run for 10 years and longer. Token consumption is still growing quickly. Power gaps make construction physically necessary. But compute demand is shifting. It is moving from “training first” to “inference-led,” and from “stacking parameters” to “competing on efficiency.”

JPMorgan analysts gave a specific judgment: inference gross margins are now 60% to 80%. If a model vendor sells AI tokens, each GW of compute can generate $20 billion to $40 billion in annual revenue. That is far above about $10 billion per GW in 2025. “Selling compute” makes money. “Buying compute” is also a good business.