AI’s Valuation Reset: Why Agent Demand Deserves a Second Look

OpenRouter token use is almost 25× January, while AI stocks have reset sharply. A source-backed look at agent demand, hyperscaler capex, credit risk and a ten-stock AI-infrastructure watchlist.

20 November 2026 · AI infrastructure · Research watchlist, not a trade instruction AI stocks are being repriced as investors question hyperscaler capital expenditure , financing costs and the return on AI investment. That scrutiny is healthy. But the demand side of the equation is developing faster than the bearish narrative suggests. The more useful question is no longer simply “how much are hyperscalers spending?” It is “how much useful work are AI systems doing, and how is that workload changing?” On that measure, the latest data are striking. The demand signal is becoming harder to ignore OpenRouter reported 790 trillion tokens in October 2026, up 37% from September and almost 25 times January’s volume . Agentic API keys generated 76% of October tokens, compared with 8% a year earlier. This is not the entire AI market. OpenRouter is one routing platform, and token counts include prompt and completion tokens. Nevertheless, the direction and magnitude provide a useful real-time signal that is independent of a chipmaker’s revenue guidance . Three changes matter: AI coding has crossed a usefulness threshold. Coding assistants have existed for years, but more capable models now complete longer, more reliable tasks rather than only suggesting snippets. More completed tasks mean more inference sessions. Agents are becoming workload multipliers. A longer-running agent does not make one request. It plans, retrieves context, calls tools, checks results and tries again. The same user objective can therefore generate many model calls. The tool ecosystem is ready for agents. MCP, coding tools, search, databases and enterprise connectors give agents more ways to perform useful work. Better tools increase the range of economically valuable tasks—and the inference required to complete them. AI coding also has a second-order effect . Every new application or AI service built with these tools can create additional users, workflows and inference demand. The initial coding session is not necessarily the end of the compute requirement; it can be the beginning of a new production workload. Intelligence is becoming a metered input For the first time, organizations can purchase machine intelligence as an on-demand input measured in tokens and compute. That changes the demand question. There is no obvious fixed inventory of tasks: when intelligence becomes cheaper, faster or more capable, users can apply it to problems that were previously uneconomic. Higher-quality output can itself require more resources. A system can reason for longer, retrieve more context, verify its answer, simulate alternatives or run multiple agents in parallel. A user may also choose a better answer rather than simply buying the same answer more cheaply. This is why lower unit prices do not necessarily lower total spending; they can unlock enough new usage to expand the market. The claim is not that demand for intelligence is literally unlimited or that exponential growth is guaranteed. Budgets and economic returns still matter. The claim is that useful intelligence has no obvious near-term saturation point comparable with a mature commodity, and the early usage data are consistent with an adoption curve driven by expanding capability rather than a one-off technology trial. Why memory is a central constraint Memory is more than a peripheral component. HBM capacity and bandwidth help accelerators handle larger models, longer contexts, bigger key-value caches and more concurrent inference sessions. Additional memory can improve throughput and enable workloads that would otherwise be slow, unavailable or require more fragmented infrastructure. Better utilization can improve the economics of the entire accelerator deployment. Memory does not automatically improve answer quality by itself; model design, training and inference methods still determine capability. But memory can remove practical constraints on deploying the most capable models and serving them efficiently at scale. That is why MU and SKHY are attractive candidates in this thesis. They sit at a potential supply bottleneck rather than relying only on which application wins. If demand for capable intelligence continues to outpace HBM and advanced-memory supply, pricing power can remain with suppliers. The risk is the reverse: capacity additions, yields or demand disappointment can eventually turn scarcity back into a memory-cycle downturn. The operators have better information than outside commentators The strongest demand argument may be the actions of the cloud providers themselves. When major US and Chinese hyperscalers independently conclude that they need more compute, they are drawing on information outsiders do not possess: current utilization, customer negotiations, order books, product roadmaps and experience with internal models that may be several generations ahead of public systems. That is a fundamentally different evidence base from an outside commentator extrapolating a recent share-price decline. These companies see both the demand and the technical direction. Their capex decisions are consequential commitments, not casual opinions. If a broad group of sophisticated competitors reaches similar conclusions despite different business models and domestic market conditions, dismissing that evidence requires a strong counterargument. There is also an important distinction between questioning the operators’ technical judgment and questioning the market’s investment logic. It is entirely consistent to respect the operators’ superior information while believing investors are overweighting near-term capex, interest rates and credit headlines relative to the future usefulness of AI. The investor’s remaining question is not “are these companies foolish?” It is “what return will this investment earn, when will it arrive, and what price is already assumed in the stock?” Operators can rationally accept lower near-term returns to defend a strategic platform. Shared competitive pressure can also produce overcapacity even when every technical forecast proves correct. Demand, capability and shareholder return are connected, but they are not identical. The appropriate response is therefore humility about outside demand forecasts—not abandonment of valuation discipline. Insider capex commitments should raise our confidence that real demand and technical progress exist; the investment decision still requires monetization, timing, balance-sheet capacity and an acceptable entry price. Token growth is not the same as profit growth This distinction is essential. OpenRouter also reported that effective prices per million tokens had fallen 57% for closed-weight models and 70% for open-weight models since January. Caching, model mix and competition are reducing unit costs. Model providers and cloud providers sit in different economic layers A model provider’s price per token includes much more than raw compute: model development, training, research, safety, software and support . A cloud provider can instead monetize accelerator time, compute units, networking, storage and related infrastructure. A reduction in an LLM provider’s token price therefore does not mechanically require the same reduction in the cloud provider’s charge per compute unit. That distinction matters because token deflation can coexist with resilient infrastructure margins. If lower application prices stimulate enough usage, cloud providers can benefit from higher utilization and aggregate compute demand even while model providers sacrifice some margin to compete. But cloud margins are not automatically insulated. Reserved-instance discounts, customer bargaining, depreciation schedules, energy costs, competition and utilization still matter. Hyperscalers that also sell first-party models or token-priced managed inference participate in both layers and may choose to subsidize one to support the other. The correct test is therefore not token price alone, but revenue per compute unit, utilization, infrastructure cost and cash return on deployed capacity. Falling prices can stimulate demand while limiting revenue per unit of compute. Investors therefore need to connect token growth to: paid utilization and customer retention; accelerator and memory utilization; cloud revenue and backlog ; gross margin after inference costs ; and free cash flow after capital expenditure. The bullish case is not “tokens rise, therefore every AI stock wins.” It is that rapidly growing useful work makes the infrastructure demand cycle more durable than a purely speculative interpretation implies. Capex is a rational response—but not a free pass If demand is growing faster than available capacity, higher capex is a normal investment response. J.P. Morgan noted that five US hyperscalers’ estimated 2026 capex had reached $697 billion, while cautioning that capex had risen from 33% of their operating cash flow in 2023 to an estimated 93% in 2026. The same analysis found that higher capex was rewarded only when accompanied by higher revenue estimates. That is the right discipline. Credit concerns also deserve more than a dismissal. Higher rates, debt issuance and long-dated commitments can reduce equity value even when end demand is real. Financing structure, customer concentration and circular vendor arrangements still need scrutiny. Anthropic’s IPO demonstrates substantial equity-market appetite: it priced at $83 per share, above its indicated range, and closed its first session at $97.17. But the same session saw the Nasdaq fall 2.3%. A successful IPO is evidence of access to capital, not evidence that credit risk has disappeared. The reset is creating a more selective entry window The correction has been meaningful. Based on unadjusted 2026 monthly highs and 19 November closes from stockfinancedata: | Company | Ticker | 2026 high | 19 Nov close | Price drawdown | |---|---|---:|---:|---:| | Micron | MU | $1,255.00 | $690.44 | 45% | | AMD | AMD | $658.52 | $489.19 | 26% | | Tencent | 0700.HK | HK$639.00 | HK$348.40 | 45% | | Nvidia | NVDA | $243.37 | $145.98 | 40% | | Broadcom | AVGO | $495.00 | $346.65 | 30% | These are price drawdowns, not claims that each company’s earnings multiple contracted by the same percentage. Entry attractiveness still depends on earnings durability and valuation against realistic forward assumptions. A research watchlist across the AI stack The original candidates—MU, SKHY, AMD and Tencent—cover memory, accelerators and an Asian application/cloud platform. Adding other bottlenecks can reduce dependence on one product cycle. | Company | Ticker | Role in the thesis | Principal issue to monitor | |---|---|---|---| | Micron | MU | HBM and memory pricing | Cyclicality, capacity additions and contract conversion | | SK hynix | SKHY | HBM scale and advanced memory | ADR premium/discount, FX, Korea listing linkage and capex | | AMD | AMD | Alternative AI accelerators and CPUs | Execution, software adoption and current China allegations | | Tencent | 0700.HK | AI applications, advertising and cloud | Monetization versus AI cost growth; China regulation | | Nvidia | NVDA | Merchant accelerators and systems | Valuation, competition, governance and customer concentration | | Broadcom | AVGO | Custom accelerators and networking | Hyperscaler concentration and program timing | | TSMC | TSM | Leading-edge manufacturing and packaging | Geopolitics, capacity economics and node execution | | Arista | ANET | Ethernet fabrics connecting AI clusters | Cloud concentration and networking competition | | Vertiv | VRT | Power delivery and thermal management | Backlog conversion, capacity and valuation sensitivity | There are already fundamental signals behind parts of this watchlist. Broadcom reported Q3 AI semiconductor revenue growth of 221%. Tencent’s Q3 release said business-services growth was driven by a 35% increase in cloud-services revenue, underpinned by AI demand. Those observations are more useful than treating “AI” as one undifferentiated trade. AMD requires particular caution after the latest short-seller report concerning China-related exposure. The allegations should be investigated, not treated as a regulatory finding—or ignored because the broader thesis is attractive. What the Oscar CCI debate added An Oscar AI Committee review of AMD reached a deliberately selective conclusion. It accepted that forward earnings are more informative than trailing P/E during an earnings inflection, but rejected any direct leap from token growth to AMD profits. Nvidia’s software ecosystem and hyperscaler custom silicon may capture a large share of incremental inference before AMD does. The committee also highlighted a portfolio-construction point: MU and SKHY are largely the same memory-cycle exposure, while AMD adds a different accelerator-execution risk and Tencent adds an application/cloud and China-regulatory vector. AMD is therefore a diversifier within this watchlist, not automatically the best standalone holding. Finally, CCI treated the China report, related legal investigations and executive selling as live risks requiring review. Its conclusion was cautious and staged—not an unconditional endorsement. That is the appropriate approach to the wider reset as well. What would invalidate the opportunity thesis? The entry case weakens if: agent-token growth slows materially for several months without a seasonal or pricing explanation; falling inference prices overwhelm volume growth and compress supplier economics; cloud customers cut orders, delay delivery or reduce backlog; memory and accelerator supply outpaces demand, leading to pricing pressure; capex rises without corresponding revenue, utilization or free-cash-flow improvement; or credit spreads and financing costs continue rising faster than operating cash generation. These are observable tests. They make the thesis falsifiable rather than dependent on enthusiasm for AI. Bottom line The market’s concern about capex and credit is rational. Treating those concerns as proof that AI demand is weak is a different—and increasingly questionable—leap. The combination of 25-fold token growth, agentic workload dominance, improving coding capability and a maturing tool ecosystem supports a stronger demand interpretation. When those data are combined with informed operators across the US and China committing their own capital to more compute, the burden increasingly falls on skeptics to explain why that evidence should be ignored. The recent reset therefore deserves serious consideration as a selective research and staged-entry opportunity, particularly across memory, networking, manufacturing and power constraints. It is not a signal to buy every AI-related stock indiscriminately. The opportunity is to use lower prices while demanding evidence that usage becomes profitable, durable cash flow . Sources and method OpenRouter October 2026 data recap OpenRouter: The Token State of Play J.P. Morgan Asset Management: Technology and AI outlook Anthropic IPO announcement Investopedia: 19 November market recap Tencent Q3 2026 results Broadcom Q3 FY2026 results Historical unadjusted prices: stockfinancedata, monthly observations through 19 November 2026. Drawdowns calculated from each series’ 2026 monthly high to its 19 November close. This article is for research and education, not personalized investment advice. The watchlist has no implied weights, price targets or assurance of returns. Additional watchlist inclusion: Cloudflare (NET) Cloudflare extends the portfolio to edge compute, inference delivery and the developer platform through which AI applications reach users. It complements memory, accelerators and data-center infrastructure with an application-delivery layer. The key monitoring points are usage monetization, infrastructure utilization, capex and the premium valuation attached to that growth.