Lu Zhang on the Signals and Noise in the AI Investment Boom

Stress testing the AI thesis

The market has delivered a run of real tests in the final weeks of Q3. Most recently, AI-linked stocks sold off alongside a broader risk-off move in mid-September before reversing as oil eased and bond yields fell.

Sell-offs have recurring narratives including doubts over whether AI spending reflects real demand or circular vendor financing, warnings from AI’s own leaders that development itself may need to slow and analysts flagging historical bubble-peak signals.

Hyperscaler Q2 earnings, while beating estimates on the surface, raised some red flags. Capital spending grew roughly three to four times faster than revenue last quarter for Microsoft (NASDAQ:MSFT), Alphabet (NASDAQ:GOOGL), Amazon (NASDAQ:AMZN) and Meta Platforms (NASDAQ:META). Free cash flow also fell - Meta’s collapsed by more than 90 percent, Amazon and Oracle (NYSE:ORCL) both went negative, with Oracle falling by close to US$24 billion.

NVIDIA (NASDAQ:NVDA) posted real profit, but its own free cash flow still fell by US$27 billion in a single quarter, largely because it’s extending credit to its own customers to help them buy its chips, while Amazon’s profit beat came mostly from a paper markup on its Anthropic stake, not the underlying business.

Meanwhile, Nvidia’s compute guarantee backing OpenAI’s data-center buildout has reportedly shrunk from an initial US$250 billion to US$105 billion, and OpenAI’s own CFO has acknowledged that Nvidia’s investment effectively comes back to Nvidia as chip purchases.

Real demand, or financing dressed up as demand?

According to Zhang, big tech is “very determined” to keep building, willing to fund it with debt and equity rather than cash because losing access to compute to a competitor isn’t a risk they can afford to take.

“They have to invest, regardless (of whether) they like it or not,” she said. Hyperscalers’ own stated logic for spending this aggressively isn’t about near-term unit economics. “They want to have control over their AI strategy and don’t want the bottleneck controlled by their competitor.”

Zhang, an early-stage investor who backs enterprise and industrial AI companies, brings a simple filter to the current wave of AI infrastructure deals.

• Whether a company has curated its own high-quality, industry-specific data.
• Whether a company has optimized its architecture to bring down the cost of running the model, not just training it.
• Whether a company already has real partnerships with the industry players who control the data and workflows it needs.

Miss any of those, in her view, and a “differentiator” is just a temporary lead.

Zhang points to budget allocation from banks, insurers and healthcare systems as the clearest outside-tech signal that demand is durable rather than manufactured. She also discounts headline revenue on its own; in her view, a seller’s market can inflate revenue without matching real deployment.

“We also want to see a very healthy growth of the unique selling number across the industry that people are actually digesting, buying and deploying more GPU and CPU. I expected we would see some back and forth,” she said, noting that GPU demand shifted visibly between early and late 2025.

On the capital markets side, she expects a “more creative approach of using different financial vehicles” to fund AI infrastructure, sees “definitely some of the bubble issue” in current valuations and calls backlog figures “very optimistic projections.”

The real risk to the AI investment thesis, in her view, could instead be a governance failure serious enough to force a pause in deployment.

The cost of getting compliance wrong

Meanwhile, the governance question is especially pressing since OpenAI disclosed that its own AI agents breached developer platform Hugging Face over several months this summer, reaching administrator-level access before the company caught it.

OpenAI has since paused reinforcement learning on its own largest planned frontier run, after preliminary evidence that its next model could cross the critical cybersecurity capability threshold in its own safety framework. The delay has already cost significant time and research resources, according to the company.

This week, at the UN General Assembly, more than 20 countries jointly proposed an international body for AI safety standards and incident reporting. Washington separately floated a bilateral hotline with Beijing for flagging AI incidents ahead of Chinese President Xi Jinping’s planned visit to Washington.

Zhang described AI adoption happening in two phases: automation first, which cuts costs without necessarily showing up in revenue, followed by optimization, when redesigned workflows finally move the needle on earnings.

A heavier compliance and reporting regime doesn’t kill that second phase, but it adds a cost line and a delay, especially for regulated industries meant to be the proof of durable demand.

The bottom line

For now, the AI buildout sits in an awkward middle ground. The demand signal is deafening, but the financing story is messier than the headlines suggest.

Fusion Fund’s framework pushes investors to look past who is training the largest model or spending the most. Lu encourages market participants to tighten their filters, focusing on companies with real data moats, architectures built for low‑cost deployment and partnerships that prove their software is a must‑have, not a demo.

As the industry weathers real stress tests, investors should treat AI not as a monolith, but as a set of sharply different bets on who will actually turn compute into cash.