AI demand called unlimited, but chip stocks fall as spending questions mount across industry

AI Demand Remains Strong While Markets Signal Caution

AI demand is "almost unlimited," according to Pat Gelsinger, former Intel CEO and now general partner at Playground Global, who identifies energy availability as "the only real limiter" to growth 2

2

. This sentiment echoes across the industry, with Lumentum reporting its photonics and optical products for data center connectivity are sold out for the next five years 2

2

. Yet AI chip stocks have experienced sharp volatility, with TSMC shares falling 4 percent despite posting record second-quarter revenue exceeding $40 billion 5

5

. The disconnect between robust AI demand and market skepticism highlights mounting concerns about the financial sustainability of the AI industry.

The market's reaction stems from TSMC raising its capital expenditure forecast to $60-64 billion, up from $52-56 billion 5

5

. This adjustment, combined with broader AI infrastructure spending projected to reach $1.2 trillion in 2027 across five major hyperscaler companies, is testing investor patience 1

1

3

3

. Samsung similarly forecast a gigantic profit rise yet saw its stock decline after a 360 percent rally over 12 months 2

2

. The PHLX chip index has gained roughly 60 percent year-to-date, pricing in years of flawless execution that leaves little room for disappointment 4

4

.

Shift from Tokenmaxxing to Valuemaxxing Reshapes AI Spending

A fundamental transformation in how enterprises approach AI compute is reshaping the industry. Companies are abandoning "tokenmaxxing"—encouraging unlimited AI usage regardless of outcomes—in favor of valuemaxxing, which prioritizes return on investment 2

2

. This shift began when Microsoft-owned GitHub switched to usage-based pricing in April, marking what BCA Research calls the "end of the token subsidy" that underwrote AI's rapid adoption 1

1

. The move toward consumption-based pricing has driven rapid growth in usage of open-source models from Chinese providers like DeepSeek and Alibaba, which offer lower costs than frontier models from OpenAI and Anthropic 1

1

2

2

.

This cost consciousness raises critical questions about AI monetization. Marc Boroditsky, chief revenue officer at Nebius, emphasizes that tokenmaxxing only makes sense when organizations see a return on investment, stating that "the CFO bringing the hammer down and slowing spend should actually be looking for value" 2

2

. The trend toward lower-cost models is pulling down spending per token, even as total usage increases 1

1

. Some analysts see this as evidence of Jevons paradox, where increased efficiency drives higher total consumption, while others fear commoditization will make AI more like electricity than software, with economic value accruing to application builders rather than model makers 1

1

.

Bond Issuance Funds AI Data Centers as Free Cash Flow Declines

Major tech companies are increasingly turning to bond issuance to finance AI infrastructure as AI spending outpaces their ability to self-fund. Oracle's aggressive borrowing for AI data centers led S&P Global to downgrade its debt rating to just one notch above junk bond status on July 9, citing deteriorating finances 3

3

. The company's financial troubles reflect a broader pattern: according to Bank of America, total capital expenditure for Oracle, Alphabet, Microsoft, Amazon, and Meta now exceeds their free cash flow—the money their businesses generate beyond operational and investment needs 3

3

.

The ratio of capex to revenue is mounting across hyperscaler companies, transforming previously capital-light tech firms into entities as capital-intensive as fossil fuel giants like Exxon Mobil and Chevron 1

1

3

3

. While Microsoft maintains a Triple-A credit rating—better than the U.S. government's—its continued AI spending spree may test that status 3

3

. Meta's announcement that it would sell excess AI compute capacity sparked market concerns about potential overcapacity, though executives like Andrew Feldman of Cerebras Systems call such cases "unique" and maintain that "demand for compute far outstrips available capacity" 2

2

.

Energy Constraints and Overvaluation Concerns Loom Large

The AI investment bubble faces two critical constraints that no amount of capital can immediately solve. Gelsinger's identification of energy constraints as the binding limit suggests that even if AI demand proves unlimited, the infrastructure to meet it faces physical bottlenecks that operate on timescales far longer than quarterly earnings cycles 4

4

. If energy becomes the primary constraint rather than chips, then semiconductor valuations rest on infrastructure the industry doesn't control 4

4

. Meanwhile, concerns about overvaluation intensify as the technology has yet to justify the nearly $1.6 trillion spent developing it over the past decade 5

5

.

Goldman Sachs estimates AI could take as long as 15 years to reach peak adoption—faster than the median 29 years for previous general-purpose technologies, but still a lengthy timeline that challenges current valuations 1

1

. The AI investment bubble narrative is shifting from contrarian viewpoint to mainstream concern, as even record-breaking earnings from companies like TSMC and Nvidia fail to sustain investor confidence 5

5

. Hyperscaler stock performance has stagnated over the past three months despite continued spending, while data center bottlenecks and the risk of malinvestment grow more apparent 1

1

. The fundamental question remains whether massive AI spending will generate proportional returns, or if the industry has priced in perfection that reality cannot deliver.