From TSMC's Record Earnings to the Next AI S-Curve: Infrastructure, Inference, and Selective Rotation

The AI trade is not a speculative bubble. It is a sustained, infrastructure-driven buildout, and the numbers are validating the exponential adoption curve. The clearest signal comes from TSMC, the world's most critical foundry. In 2025, the company posted record annual revenue of $122.42 billion, with AI and HPC processors accounting for 58% of sales-roughly $71 billion. That segment grew 48% year-over-year, demonstrating that demand is not just present but accelerating at a steep rate.

To meet this demand, TSMC is committing unprecedented capital. The company has raised its 2026 capital expenditure forecast to $52-$56 billion. This figure alone dwarfs the combined spending of its closest competitors, Intel and Samsung, from the previous year. More importantly, it reflects a first-principles understanding of the buildout: the lead times for new advanced fabs are about three years, so planning must be done now. As CEO C.C. Wei stated, the sheer scale of this investment-between $52 and $56 billion-makes him "very nervous," but also confirms he has double-checked the demand with cloud service providers and found it to be real and healthy.

This pattern of underestimation is systemic. Analyst consensus has consistently failed to capture the true pace of AI capital spending. As Goldman Sachs Research notes, analyst estimates have consistently underestimated capex spending related to AI, with real spending exceeding estimates by over 50% in both 2024 and 2025. The market is now catching up, with the consensus estimate for 2026 AI hyperscaler capex climbing to $527 billion. The key point is that this isn't a one-time surge; it's the validation of an exponential curve where infrastructure spending is outpacing even the most aggressive forecasts.

This sets the stage for a critical market rotation. The initial phase rewarded all AI infrastructure companies for their big bets. But as the Goldman report details, investors have rotated away from AI infrastructure companies where growth in operating earnings is under pressure and capex is being funded via debt. The next phase will favor those with a clear link between capital investment and future revenue, moving beyond the pure capex story to the productivity gains and platform efficiencies that will follow. The exponential buildout is real; the market is now sorting the builders from the borrowers.

The Shifting Paradigm: Inference as the New Infrastructure Layer

The AI S-curve is entering a new phase, and the shift to inference is the defining work of 2026. The market is moving away from the massive, capital-intensive training of foundational models and toward the constant, scalable use of those models to answer questions and perform tasks. This isn't a retreat from compute demand; it's a fundamental reconfiguration of it. According to Deloitte, inference will drive roughly two-thirds of all compute this year, up from half in 2025. This is the new infrastructure layer being built on top of the existing one.

The critical implication is that this shift will fuel a dedicated market for inference-optimized chips, projected to exceed $50 billion in 2026. This creates a clear bifurcation in the chip market. While inference chips will grow rapidly, the heavy lifting for complex model operations will still be performed on cutting-edge, power-hungry AI chips worth $200 billion or more. In other words, the ecosystem is moving toward a "both-and" architecture, not an "either-or" one.

Yet, despite the efficiency gains in inference chips, the overall compute demand will keep climbing. Deloitte's key point is that enterprises will need all the data centers and enterprise on-premise AI factories currently being planned. The reason is twofold. First, inference workloads are not simple. As models become more complex and are used in advanced techniques like "post-training" and "test-time" scaling, the per-query compute cost can skyrocket. Second, the sheer volume of inference requests from billions of users and devices will add up to a massive new load. In practice, this means the exponential buildout of data center capacity and electricity, already validated by TSMC's capex, is far from over. The market is not building fewer data centers; it is building a new, specialized layer of inference infrastructure on top of the existing, massive compute base.

Market Rotation and Selective Valuation: Who Gets the Exponential Growth?

The market's initial broad bet on AI infrastructure is giving way to a sharper, more selective rotation. Investors are no longer willing to reward all big spenders equally. The clear signal is that capital is flowing away from AI infrastructure companies where growth in operating earnings is under pressure and capex is being funded via debt. This divergence has been stark, with the average stock price correlation across large public AI hyperscalers falling from 80% to just 20% since June. The logic is simple: the exponential buildout is real, but the market is now sorting the builders from the borrowers.

Goldman Sachs Research frames the next phase of the trade accordingly. The focus will shift from pure capex stories to AI platform stocks and productivity beneficiaries. The expectation is that the next wave of gains will come from companies demonstrating a clear link between massive capital investment and future revenue, moving beyond the initial infrastructure buildout to the platform efficiencies and economic productivity that will follow. This is a pivot from funding the rails to monetizing the trains.

This selective approach is playing out in sharp market signals. On one side, AMD shares have surged 11% this week, with the company reportedly almost sold out for 2026. The rally is built on tangible demand, with the company considering price increases and analysts calling it a top pick. On the other side, a major institutional player is taking the opposite view. Stanley Druckenmiller sold his entire Nvidia position, a move that adds weight to the narrative that the stock's easy gains may be behind it. While Nvidia remains a leader, its recent performance has been muted, rising less than 4% over the past three months.

The bottom line is a market recalibrating its view of exponential growth. The infrastructure S-curve is validated, but valuation is now tied to sustainability. The rotation shows investors are demanding proof of a path from capital expenditure to durable earnings. They are rotating into companies with strong cash flows and clear monetization strategies, while stepping back from those where the financials are stretched. This is not a rejection of AI, but a maturation of the trade.

Catalysts, Risks, and What to Watch

The thesis of sustained exponential infrastructure growth now faces its first real tests. The market has validated the buildout, but the next phase requires confirmation that this demand is robust and sustainable, not masked by accounting or policy shifts.

The immediate watch is on upcoming tech earnings. As Swissquote analyst Ipek Ozkardeskaya noted, upcoming tech results will be watched for any signs of accounting sleights of hand. This caution is warranted. With companies reporting record AI-driven revenue, the risk of aggressive revenue recognition or cost deferrals to meet lofty expectations increases. The market will scrutinize these reports not just for top-line growth, but for the quality of earnings and the durability of the underlying demand. A single earnings miss or a warning about inventory could quickly challenge the narrative.

More broadly, the pace of hyperscaler capex announcements remains the critical signal. The history here is clear: analyst consensus has consistently underestimated the true scale of investment. As Goldman Sachs Research highlighted, analyst estimates have consistently underestimated capex spending related to AI, with real spending exceeding estimates by over 50% in both 2024 and 2025. The market is now pricing in a consensus of $527 billion for 2026 AI hyperscaler capex. Any new announcements that significantly exceed this figure would be a powerful vote of confidence in the exponential curve. Conversely, a slowdown or a shift in funding sources would be a major red flag.

Two key risks could pressure this growth story. Geopolitically, the recent U.S.-Taiwan trade deal, while aimed at boosting investment, introduces new variables. The agreement's details on tariffs and investment commitments will be watched for any unintended friction in the critical semiconductor supply chain. More pressing for the market is monetary policy. Recent economic data has supported expectations of no immediate Fed rate cut. Lower-than-expected U.S. jobless claims data Thursday, paired with strong Philadelphia Fed and Empire manufacturing surveys, supported expectations that a further Federal Reserve rate cut is unlikely in the immediate term. This is a double-edged sword. It provides stability but also removes a tailwind for growth stocks, which have thrived on cheap capital. Any shift in the Fed's stance, or a surprise in economic data, could quickly alter the risk appetite for high-multiple infrastructure plays.

The bottom line is that the market is moving from validating the S-curve to testing its slope. Investors must watch for accounting transparency, hyperscaler spending confirmations, and the twin pressures of geopolitics and monetary policy. The exponential buildout is real, but its path will be confirmed by the numbers, not the narrative.