AI Hardware Profits Soar While Software Valuations Get Cut in Half: The Truth Behind Institutional Voting

Fire and Ice: An AI Industry Divergence Unfolding in Real Time

As 2025 passes its midpoint, the AI industry presents a thought-provoking picture: on one side, NVIDIA's market cap has surpassed $3 trillion, TSMC's orders are booked out two years, and Broadcom's custom chip revenue has doubled. On the other side, a large number of AI SaaS companies have seen their valuations shrink 40%-60% from peak levels, and several high-profile AI startups are facing funding difficulties or even layoffs and downsizing.

Institutional investors are making their choices with real money — going heavy on hardware and reducing software positions. Behind this vote with their wallets lies a deeper logic: the AI industry is transitioning from concept-driven hype to a value-driven reality.

The Hardware Side: A 'Money Printing Machine' of Certain Profits

The fundamental reason hardware manufacturers are favored by institutions comes down to the high certainty of their profits.

NVIDIA remains the biggest winner at this feast. In fiscal year 2025, its data center business revenue grew over 100% year-over-year, with gross margins consistently above 70%. Regardless of whether downstream AI applications have truly validated their business models, as long as companies are still training models and expanding clusters, GPU demand remains rigid. Just as the people who profited most during the Gold Rush were often those selling shovels, NVIDIA plays exactly that role.

TSMC has also benefited enormously. Advanced process capacity is in short supply, with 3nm and the soon-to-be mass-produced 2nm processes almost entirely locked up by AI chip orders. TSMC's Q1 2025 revenue grew over 35% year-over-year, with the share of AI-related revenue continuing to climb.

Additionally, custom chip (ASIC) manufacturers such as Broadcom and Marvell are also experiencing an explosive growth phase. Demand for Google TPUs, Amazon Trainium, and Meta's in-house chips has filled their order books to unprecedented levels. HBM (High Bandwidth Memory) suppliers like SK Hynix and Samsung are likewise raking in massive profits.

The institutional logic is simple: the hardware side has real revenue growth, verifiable profit data, and continuously expanding order backlogs. In a market permeated with uncertainty, certainty itself commands the greatest premium.

The Software Side: Squeezing Water Out of the Valuation Bubble

In stark contrast to the hardware side's prosperity, the AI software sector is experiencing a collective cooldown.

The most significant valuation corrections are among AI SaaS companies. From 2023 through early 2024, a large number of AI software companies achieved valuation multiples far exceeding their fundamentals on the narrative of "AI empowerment." Some companies saw their PS (Price-to-Sales) ratios reach as high as 50x or even 100x, but as the market gradually returns to rationality, these valuations are being ruthlessly compressed.

The core issues exist on three levels:

First, business models remain unproven. Most AI software companies face an awkward reality: users are willing to try, but unwilling to keep paying. The funnel from free trial to paid conversion is extremely steep, and customer retention rates are generally falling short of expectations. Many companies' ARR (Annual Recurring Revenue) growth has rapidly decelerated after an initial burst.

Second, moats are paper-thin. As large model capabilities trend toward homogeneity, AI applications built on API calls inherently lack barriers to entry. An AI writing tool or an AI customer service chatbot can be replicated by competitors with highly similar functionality within weeks. Low switching costs mean extremely low user loyalty, making price wars virtually inevitable.

Third, compute costs devour profits. The gross margin structure of AI software companies is fundamentally different from traditional SaaS. Traditional SaaS typically enjoys gross margins of 70%-80%, while AI-native applications, due to heavy reliance on GPU inference resources, often see gross margins of only 50%-60%, with some companies even lower. This means that even with revenue growth, profit margins are severely squeezed by compute costs.

Even star companies have not been spared. According to multiple media reports, several AI unicorns once valued at billions of dollars have encountered "down rounds" in their latest funding, with investors beginning to demand clear paths to profitability rather than mere user growth numbers.

The Deeper Logic Behind Institutional 'Voting with Their Feet'

The migration of institutional capital from software to hardware fundamentally reflects the market's repricing of the AI industry value chain.

Logic One: Certainty of profits takes priority over the imagination of growth. In a high-interest-rate environment, the opportunity cost of capital has risen significantly. Institutions are no longer willing to pay for distant profitability expectations and instead prefer to allocate to targets that are "making money right now." NVIDIA delivers above-expectation profits every quarter, while most AI software companies cannot even provide a timeline for breakeven.

Logic Two: Value distribution across the industrial chain is severely uneven. Currently, AI industry profits are highly concentrated in the upstream infrastructure layer — chips, servers, and data centers. The midstream model layer and downstream application layer are either burning cash for scale or struggling to survive in a red ocean. This "top-heavy, bottom-light" profit structure is unlikely to change in the short term.

Logic Three: The 'killer application' for AI has yet to emerge. Although ChatGPT ignited public attention, truly transformative AI applications that can replace existing workflows and create enormous incremental value remain scarce. Enterprise AI procurement budget growth has been slow, with many still in the "pilot" stage rather than "full deployment." Without scaled application deployment, the software side will struggle to deliver on its growth promises.

Echoes of History: Lessons from the Dot-Com Bubble

This scene is not without precedent. Before and after the dot-com bubble burst in 2000, hardware infrastructure companies like Cisco and Sun Microsystems were among the most profitable targets, while countless ".com" companies vanished into thin air. But the second half of that story is equally worth noting — the companies that ultimately created enormous value were application-layer giants like Google, Amazon, and Facebook, which rose from the ashes.

The AI industry may be following a similar script: hardware certainty is strongest in the short term, but in the long run, once a "killer application" emerges and validates its business model, the value unleashed by the application layer will far exceed the infrastructure layer.

The key question is — how long is this 'long term,' exactly?

When Will the Turning Point Arrive?

For the software side to regain institutional favor, several key conditions must be met:

1. Gross margin improvement: As inference costs continue to decline (through model distillation, inference optimization, and the proliferation of specialized chips), the gross margins of AI applications are expected to gradually converge with traditional SaaS. This is the prerequisite for valuation reassessment.

2. Establishment of differentiated moats: AI application companies with unique data assets, deep industry know-how, or strong network effects will gradually distance themselves from "wrapper" products.

3. Substantive growth in enterprise AI spending: Only when AI transitions from "experimental budgets" to "core IT budgets" can application-layer revenue growth gain sustainable support.

4. Emergence of killer applications: Similar to what the iPhone did for mobile internet, the AI space needs a benchmark product that truly changes user behavior and creates irreplaceable value.

Multiple analysts predict this turning point may arrive between 2026 and 2027. By then, inference costs will have dropped to less than one-tenth of current levels, AI Agent technology will have matured, and enterprise application scenarios will become much clearer.

Conclusion: From 'Storytelling' to 'Reading Balance Sheets'

The core logic of the AI industry is shifting from "storytelling" to "reading balance sheets." This is not a bad thing — the squeezing out of bubbles is precisely what marks the maturation of an industry.