01 Trigger Event

On November 11, 2026, Bloomberg aired an interview with Citi’s Olaolu Aganga. Her core assessment had three elements: strong earnings and attractive valuations continue to support U.S. equities; AI remains in a long-term investment cycle whose productivity gains have yet to be fully realized; and inflation could surprise to the upside, pushing interest rates higher and relatively favoring short-duration fixed income and gold.

This was a macro strategy judgment, not a new model release or a change in model API pricing. Its value lies in placing the AI investment cycle within the macro constraints of interest rates, asset duration, and capital expenditure, rather than discussing model capabilities alone.

02 What It Really Means

The issue is not whether AI will continue to improve productivity, but how the cost of capital evolves before those productivity gains materialize.

Aganga’s framework ties the AI investment cycle to the inflation outlook. AI infrastructure investment, enterprise adoption, and the resulting productivity improvements take time, but the waiting period is not free. If inflation runs hotter, markets will demand higher rates, and the low-cost capital on which AI projects might otherwise rely may no longer be available.

This is also an Andrew Grove-style question: not simply whether AI is a long-term trend, but which companies have enough cash flow and execution capability to withstand economic pressure until the inflection point arrives.

For AI builders, this assessment matters more than the claim that the AI investment cycle still has a long runway, because the latter easily becomes an unfalsifiable narrative. The key variables to track are:

  • whether hyperscalers continue to raise capex forecasts;
  • whether falling inference prices translate into incremental demand as GPU supply expands;
  • whether enterprise customers move from experimental budgets to sustained deployments; and
  • whether higher rates worsen the runway of financing-dependent AI startups.

I have not run Citi’s macro model internally and cannot confirm whether Aganga’s inflation and gold allocation calls will materialize. But she identifies a structure that cannot be ignored: demand for AI may continue to grow, yet the valuation of AI assets may not expand in lockstep.

03 Historical Analogy / Structural Comparison

The current situation resembles the internet investment cycle of the late 1990s more than the 2007 iPhone launch itself.

The 2007 iPhone established a new platform for human-computer interaction and mobile computing. Over the following years, telecom operators and device manufacturers converted capital expenditure into network construction, while applications and cloud services eventually absorbed demand. Capital markets priced future demand first, and infrastructure investment preceded stable cash flows.

The current AI cycle follows a similar sequence: Anthropic, OpenAI, and Google are competing for models and compute capacity; NVIDIA is capturing capital expenditure; cloud providers are building GPU clusters; and developer tooling is packaging model capabilities into billable workflows.

The similarities, however, end at “build capacity first, wait for demand later.” A critical difference is that AI inference costs and token prices continue to fall while model capabilities diffuse. Capital expenditure may therefore be constrained by competition rather than producing the stable physical bottlenecks seen in traditional data centers.

Consequently, what investors price may not be the size of AI investment itself, but three factors:

  1. who controls distribution;
  2. who can use model capabilities to create workflows that customers are willing to renew; and
  3. who can absorb falling prices while maintaining sufficient gross margin.

The comparison with AWS in 2014 is more direct. AWS did not win every customer through any single feature; it accumulated switching costs through compute, storage, pricing, and developer familiarity. The opportunity for model APIs is similar: the gateway or platform that ultimately wins may not own the strongest model, but may offer lower routing costs, more reliable availability, and deeper developer integrations.

Nevertheless, I have not run the unit economics of individual providers internally, so I cannot confirm whether today’s token-price curve will ultimately reproduce the scale effects of cloud computing. The analogy may be overly optimistic.

04 What It Means for AI Builders

The immediate priority is not to cut AI budgets across the board, but to reassess capital intensity.

For model API consumers and AI startup founders, four statements should be reviewed first:

  • Revenue statement: whether customers are moving from pilots to production and whether AI-related revenue is genuinely entering ARR;
  • Inference cost statement: the cost shares of input, output, and cached tokens, and whether routing can adjust dynamically as model prices change;
  • Cash burn statement: if token prices keep falling and GPU capacity expands, whether the company will achieve gross-margin expansion or simply pass the savings to customers; and
  • Financing exposure statement: whether the pace of financing, customer concentration, and fixed-cost structure can still support the current runway if rates exceed expectations.

For developer tooling teams, Aganga’s observation that “AI productivity gains are still ahead” is especially worth tracking. The real validation metrics are not the number of demos, but whether AI-enabled teams deliver faster, shorten review cycles, and reduce defect rates. If productivity does not appear in these operating metrics, AI investment remains merely a capex narrative.

For model-access platforms such as opcx.ai, my judgment is that a gateway’s value cannot rest solely on aggregating more models. Model count is a replicable feature, while distribution and routing constitute the moat. The product should let customers directly observe latency, quality-adjusted cost, cache hit rate, and failure rate across models, with automatic routing by task.

There are three immediate actions worth taking:

  1. break down customer model spending by workload rather than looking only at total tokens;
  2. add prompt caching and batch API paths for highly repetitive tasks; and
  3. incorporate interest rates and cloud pricing into capacity procurement assumptions, avoiding locked-in costs if GPU prices and financing conditions deteriorate simultaneously.

I have not run customer workloads internally, so these are not assertions about any customer’s gross margin. They are, in my view, the decision corrections that matter most this week.

05 Counterargument / Risks

I may be overestimating the degree to which macro rates constrain the AI industry.

First, Aganga also acknowledges that strong earnings and valuations continue to support U.S. equities. If hyperscaler cash flow is sufficiently robust, higher rates may not reduce their capex. NVIDIA, Microsoft, Google, and Amazon could continue investing because competitive advantage requires early access to scarce capacity, not necessarily the lowest possible cost.

Second, unrealized AI productivity gains may mean the opportunity remains large, or that enterprise adoption resistance is stronger than expected. If deployment costs are high, evaluation practices remain immature, and data governance is complex, falling model prices will not automatically create demand.

Third, applying the AWS analogy to model APIs may overlook the lower switching costs at the model layer. Developers can switch providers quickly through the OpenAI SDK, Anthropic SDK, or MCP; historical invocation data, evaluations, and workflow integrations may form the real barriers. If opcx.ai does nothing beyond transparent forwarding, it could be squeezed by cloud providers or capabilities built into IDE platforms.

The most concerning combination is rising capex alongside delayed enterprise productivity gains, while falling token prices weaken revenue growth and higher rates increase financing costs. That would mark the inflection point at which the AI investment cycle shifts from waiting for returns to returns falling short of expectations.

My conclusion is therefore not that the AI cycle is over. Rather, the industry’s long-term trend may remain intact, but returns will diverge sharply across stocks, infrastructure companies, and application startups. The question is not whether AI has a future, but who bears the cost of capital, who controls distribution, and who can convert lower prices into greater demand during this cycle.