AI’s Great Divergence: Demand Is Exploding While Markets Panic

AI stocks collapsed as agent demand accelerated. A source-backed review of Nvidia, token deflation, compute economics and the wider infrastructure opportunity.

21 November 2026 · AI infrastructure · Research watchlist, not a trade instruction Friday’s selloff looked like a verdict on AI. The Nasdaq Composite fell 5.28%, Nvidia dropped 14%, AMD fell 13%, and semiconductor and memory ETFs declined 11% and 13%, respectively. The S&P 500 fell for a seventh consecutive session. Yet the same week brought Nvidia’s latest earnings and Anthropic’s successful IPO . Nvidia reported quarterly revenue of $115.8 billion, up 103% year over year, and its latest platform data pointed to continued infrastructure demand. This is the central contradiction investors need to resolve: is the market discovering weak demand, or is it repricing strong demand because of valuation, financing and company-specific risk? The earlier AI valuation-reset thesis argued for the second interpretation. The latest Nvidia analysis from Beth Kindig’s I/O Fund adds five important arguments: demand exceeds supply, cheaper tokens can increase compute consumption, the market is broadening beyond accelerators and hyperscalers, agentic AI requires a new infrastructure stack, and Nvidia’s valuation has compressed alongside that growth. Those arguments deserve serious consideration—but they are not interchangeable with proof of future stock returns. Friday’s panic is real; so are the operating results Friday’s decline was not a minor adjustment. According to Investopedia’s November 20 market recap, the Nasdaq lost 10% for the week, the S&P 500 lost 8.6%, and the VIX approached 36. Anthropic also fell approximately 9% one day after its record IPO. The market was therefore rejecting more than a single earnings report. It was reassessing the entire AI financing and valuation complex. But that makes it especially important to distinguish three layers of evidence: Observed demand: tokens processed, utilization, rental rates and delivered revenue. Future commitments: cloud backlog , supply agreements and management guidance . Equity risk: valuation, accounting questions, leadership transition, financing costs and customer concentration . The first layer has remained strong. The second is substantial but not guaranteed revenue. The third explains why strong operating results and falling share prices can coexist. Demand is increasingly contracted, not merely imagined Nvidia’s Q3 FY2027 results showed $115.8 billion of revenue and continued data-center growth. Its quarterly presentation reported that Vera Rubin added nearly $20 billion of revenue in its first production quarter. The I/O Fund’s November 20 analysis adds two larger demand indicators: cloud-industry backlog rising from $2 trillion to $2.5 trillion in three months, and Nvidia’s supply-related commitments reaching $390 billion. These commitments reflect the capacity being secured to meet expected demand, including expensive memory and related infrastructure. Contracted backlog is a stronger signal than speculative orders placed without customers. It is not, however, the same as recognized revenue or free cash flow . Investors should still examine cancellation rights, delivery schedules, customer financing and whether the commitments convert at attractive margins. This is also where the hyperscalers’ information advantage matters. Major US and Chinese operators see utilization, customer negotiations and unreleased model roadmaps that outside commentators cannot see. When many sophisticated operators independently commit capital to additional compute, their actions are evidence that demand and future model capability are real. The question for investors is not whether those operators are foolish. It is whether the timing, price and financing of the investment produce an adequate shareholder return. Token deflation can strengthen—not weaken—compute demand OpenRouter reported 790 trillion tokens in October, almost 25 times January’s volume , with agentic API keys producing 76%. At the same time, effective prices per million tokens fell 57% for closed-weight models and 70% for open-weight models since January. That apparent contradiction is the demand flywheel. Lower prices make more applications economic. More capable coding and longer-running agents then increase the number of model calls per user objective. New applications built with AI coding create further production workloads of their own. Nvidia’s earnings commentary adds an important confirmation from the infrastructure layer: despite token prices falling nearly 60% since May, the I/O Fund notes that older Hopper GPUs were recontracted at up to twice their original price, while B300 rates rose more than 60% from introduction. This supports a distinction made in the earlier thesis: LLM providers and cloud-compute providers sit in different economic layers. A model provider can cut token prices to compete, while a compute provider rents accelerator time, networking, storage and capacity. There is no mechanical requirement for compute prices to fall by the same percentage as model API prices. The boundary is not absolute. Hyperscalers also operate first-party models and managed inference , and they can subsidize one layer to support another. Utilization, depreciation, energy costs, customer discounts and competition still determine margins. But falling token prices alone do not establish falling infrastructure profitability. Intelligence is becoming a metered input—and agents multiply the workload The structural opportunity is that organizations can now purchase intelligence on demand. When intelligence becomes cheaper or more capable, it can be applied to problems that previously did not justify human or software investment. Higher-quality outcomes can require more resources rather than fewer. An agent can reason longer, retrieve larger contexts, verify its work, simulate alternatives and coordinate parallel tools. MCP and other connectors turn model output into completed tasks across real systems. This means an agentic workflow is not equivalent to a single chatbot query. It can involve dozens or hundreds of calls, tool dispatches and verification steps. The resulting infrastructure requirement extends beyond accelerator GPUs to CPUs, HBM , networking, storage, security and power. The I/O Fund highlights Nvidia’s agentic architecture as an example: Vera CPUs run agent harnesses, Rubin GPUs handle inference, and BlueField DPUs provide monitoring and policy enforcement. Whether or not Nvidia captures every layer, the broader point is important: agentic AI creates demand across a system, not just one chip category. The AI trade is broadening beyond the hyperscalers One bearish argument is that five hyperscalers cannot support an entire AI economy indefinitely. That is true. But the answer may be broader participation rather than an immediate collapse. The I/O Fund estimates that non-hyperscalers now account for 44% of Nvidia’s data-center revenue, with AI Cloud, Industrial and Enterprise revenue reaching $47 billion, up 18% sequentially. If enterprise, industrial and specialized AI-cloud demand continues to grow, slower percentage growth from the largest hyperscalers becomes less decisive. The law of large numbers also matters. Once annual capex reaches hundreds of billions of dollars, percentage growth can slow while absolute spending continues to rise. That is not necessarily a contraction in physical demand. Diversification across customers is healthier than dependence on a handful of buyers. However, it also means Nvidia’s future growth may be distributed across more suppliers and architectures. The overall AI market can expand while any individual company’s market share declines. The valuation reset changes the entry question The I/O Fund estimates Nvidia’s trailing and forward P/E ratios at approximately 15.8× and 15.3×, respectively, after the latest decline. Those are the author’s dated estimates, dependent on earnings definitions and consensus forecasts, rather than guarantees of intrinsic value . The key point is that earnings growth can compress a multiple even before the share price fully reflects it. If revenue and earnings continue growing while the price falls, the investment question changes from momentum to durability: how much of the expected growth survives competition, supply expansion and margin normalization? Oscar CCI added an important valuation caution: annualizing one exceptional quarter can make the resulting multiple look artificially precise. If earnings estimates are trimmed, the multiple rises immediately. “Cheap relative to growth” is therefore a conditional conclusion, not a permanent property of the share price. The earlier valuation-reset watchlist remains relevant because different companies occupy different bottlenecks: | Company | Ticker | Exposure | Main question | |---|---|---|---| | Nvidia | NVDA | Accelerators, systems and software | Accounting review, leadership, share and margins | | Micron | MU | HBM and advanced memory | Contract conversion and memory-cycle pricing | | SK hynix | SKHY | HBM scale and supply | Capacity, ADR/FX linkage and customer commitments | | AMD | AMD | Alternative accelerators and CPUs | Software adoption, execution and China exposure | | Tencent | 0700.HK | AI applications, advertising and cloud | Monetization versus AI cost growth | | Broadcom | AVGO | Custom accelerators and networking | Customer concentration and program timing | | TSMC | TSM | Leading-edge manufacturing and packaging | Node execution, capacity economics and geopolitics | | Arista | ANET | Ethernet fabrics for AI clusters | Cloud concentration and competition | | Vertiv | VRT | Power and thermal infrastructure | Backlog conversion, capacity and valuation | | Cloudflare | NET | Edge compute, inference delivery and developer platform | Usage monetization, utilization and premium valuation | MU and SKHY should not be treated as two independent memory bets. NET adds a different application-delivery layer, while VRT captures the physical power and cooling constraint. This is a research watchlist, not a weighted portfolio or a claim that every name offers the same entry quality. What the Oscar CCI review changed CCI accepted the demand and valuation argument but rejected any claim that this is a clean, high-conviction entry. It emphasized that strong usage and rental-market data are not themselves proof of Nvidia revenue, margin durability or correct revenue recognition. The accounting question matters precisely because it can affect the conversion of apparent demand into reported financial results. The committee also found Nvidia’s balance sheet to be a meaningful counterweight: this is not primarily a solvency story. But accounting, legal and leadership risks can still justify a lower multiple even when operating demand remains strong. Finally, CCI’s correlation analysis found NVDA, MU and SKHY moving as closely linked semiconductor exposures during the current window. Buying all three may therefore concentrate the same sector factor rather than diversify it. The wider watchlist is intended to identify different bottlenecks, but correlations still need to be checked before position sizing. What would invalidate the bullish interpretation? The thesis fails or weakens if: cloud backlog or supply commitments fail to convert into recognized revenue and cash flow ; rental rates fall because supply catches up or customers delay deployments; token-price declines stop generating enough incremental paid usage; agent adoption fails to produce durable enterprise retention and measurable productivity; memory and accelerator margins compress faster than volume grows; hyperscaler and enterprise customers simultaneously reduce orders; Nvidia’s accounting examination uncovers materially misleading financial reporting; or the CEO transition disrupts execution during a critical product ramp. These are observable tests. They prevent the demand thesis from becoming an unfalsifiable story. Bottom line The market is asking whether the AI trade is over. The operating evidence suggests a narrower answer: the easy, valuation-insensitive phase may be over, but the demand for purchasable intelligence is still expanding. Nvidia’s delivered growth, rising infrastructure commitments , resilient GPU rental rates, agentic workload growth and a broadening customer base support the view that Friday’s selloff was not evidence of weak underlying demand. The correction may therefore offer a selective research and staged-entry opportunity across the AI stack. That does not eliminate accounting, governance , financing or valuation risk. The strongest approach is neither panic nor blind confidence: use the reset to examine which companies can turn growing intelligence consumption into durable, profitable cash flow. Sources and method I/O Fund: Nvidia Stock—5 Reasons the AI Trade Has Only Begun, supplied PDF, 20 November 2026. Author-specific valuation estimates and portfolio statements are attributed to the source. Nvidia Q3 FY2027 results and quarterly presentation. OpenRouter October 2026 data recap. Investopedia: 20 November market recap. J.P. Morgan Asset Management: Technology and AI outlook. Oscar AI Committee reviewed the NVDA thesis after the 20 November selloff, including demand, accounting, leadership, margin and customer-concentration objections. This article is for research and education, not personalized investment advice. The watchlist has no implied weights, price targets or assurance of returns.