The hidden accounting mismatch powering AI and semiconductor profits

AI is driving one of the biggest capital spending booms in tech history. Chipmakers, memory suppliers, and data center builders are reporting soaring profits. Yet many of the companies actually using AI – from hyperscalers to model labs – are still burning huge amounts of cash.

Under the surface, there’s a quiet accounting mismatch that makes AI infrastructure look more profitable than it really is at the system level. At the same time, investors are trying to figure out where the real long-term value lies: in chips, data centers, or somewhere else in the stack.

The core mismatch: who books profits and who books assets

One of the most important ideas is the basic disconnect in how AI spending shows up in financial statements.

The companies selling the “picks and shovels” of the AI boom – GPUs, servers, networking gear, and power equipment – recognize revenue and profit immediately when they ship hardware. Their income statements look fantastic right now.

The buyers of that same hardware – hyperscalers like Microsoft, Google, Amazon, Oracle, and newer AI cloud providers – usually don’t expense these purchases right away. Instead, they capitalize the cost and depreciate it over years. Much of this spending sits in “construction in progress” until a data center is online and generating revenue.

The result: vendors show a profit surge today, while the customers’ true economic cost is spread out and partially hidden in future depreciation. At the system level, the AI ecosystem may be less profitable than it appears if you only look at chipmakers and equipment suppliers.

Lessons from the dot-com and telecom boom

To understand how this can end, it helps to look back at the late 1990s and early 2000s.

During the internet build-out, telecom and networking companies reported booming orders and profits as enterprises and carriers rushed to install routers, switches, and fiber. From mid-1998 to mid-2000, S&P 500 operating earnings rose about 30%.

Then order books collapsed. From 2000 to 2001, S&P earnings fell roughly 40% – about as much as during the global financial crisis – even though the recession itself was mild. Companies had overbuilt based on wildly optimistic assumptions like “internet traffic is doubling every quarter” that later turned out to be exaggerated.

The parallel to AI: today’s forecasts of near-limitless compute demand at current prices may prove too aggressive. History suggests exponential growth curves eventually hit real-world constraints – whether technological, financial, or physical.

How AI is already boosting micro-level productivity

At the company level, though, AI is clearly having an impact.

Across hundreds of “hard tech” companies – semiconductors, hardware, and related fields – operating profits have risen sharply over the past few years while headcount has barely grown or even declined. Many executives attribute this directly to AI tools improving internal efficiency.

So even if the macro impact on GDP and aggregate corporate profits is uncertain, the micro impact on individual businesses is already visible: more output and earnings with roughly the same number of people.

Why GPUs are tight and rental prices are spiking

One of the clearest signs of demand pressure is in GPU rentals.

In a normal environment, GPU rental prices should fall over time as new, more efficient architectures arrive. New chips deliver more compute per dollar, so older GPUs typically get cheaper every year.

Instead, since early 2025, rental prices have reversed and surged. Even 6–8-year-old GPUs have seen rental rates jump 40–50% as of early this year. The reason is simple: token usage – the amount of compute consumed by training and inference – is exploding faster than new supply can come online.

That tightness has dramatically improved the near-term economics for specialized AI cloud providers and GPU lessors. But it also highlights how dynamic and fragile this market is: a shift in architecture or a big supply response could flip the economics again.

Neo-clouds vs. hyperscalers: finance business, not pure tech

A growing group of companies – often called “neo-clouds” – buy GPUs from Nvidia and others, rent power and space in data centers, then lease that GPU capacity to hyperscalers and AI startups.

On paper, this looks like a high-tech business. In practice, it’s closer to an equipment leasing or finance model:

• They borrow or raise capital to buy GPUs.
• They secure data center capacity and power.
• They sign multi-year contracts to lease that capacity to larger customers.

Their returns on invested capital, even under optimistic assumptions and long GPU lifetimes (10 years in some models), often pencil out in the mid-single digits – 5–8% pre-tax. That’s surprisingly low given today’s extreme shortage and pricing power.

Meanwhile, hyperscalers could always choose to buy GPUs directly. The fact that they sometimes rent from intermediaries suggests a mix of supply constraints, timing, and vendor strategy (for example, Nvidia wanting to avoid dependence on just a few mega-buyers by supporting alternative channels).

From an investor’s perspective, this raises a key question: if the middlemen are only earning single-digit returns in the best of times, where will their economics land once the market normalizes?

Depreciation, asset life, and the timing problem

Another critical issue is how long GPUs and AI hardware will remain economically useful.

Some cautious models assume a 10-year life for GPUs, effectively running them 24/7 for a decade. That’s generous, and it helps make leasing models look better on paper. Others argue a realistic life might be closer to 5–7 years before performance and efficiency gaps make older GPUs uneconomic for cutting-edge workloads.

On top of that, there’s a timing lag:

• Hyperscalers spend heavily on chips, construction, labor, and interest.
• Those costs are capitalized as “construction in progress.”
• Depreciation only starts once the data center is live and generating revenue.

This means a lot of the true cost of today’s AI build-out hasn’t yet hit income statements. When it does, reported earnings for the big spenders could come under pressure even if revenue keeps growing.

Data centers in space and extreme compute forecasts

Some of the boldest AI visions involve building data centers in space to access abundant solar power and escape Earth’s power constraints. The logic is tied to extreme forecasts of future compute demand.

One prominent estimate suggests the world may eventually need 1 terawatt of compute capacity for AI – roughly equivalent to the entire current US power grid. By comparison, today’s AI-related data center build-out is on the order of 15 gigawatts of capacity. That’s a 60–70x gap.

However, the economics of space data centers are daunting:

• Launch costs and reliability risks are enormous.
• Cooling and radiation management in space are complex and expensive.
• Maintenance and redundancy become far harder and costlier than on Earth.

On Earth, power currently represents only about 5–7% of data center revenues. So even if space power is “free,” it doesn’t automatically make the overall economics work once you factor in everything else.

Ultimately, these extreme scenarios depend on whether current AI scaling laws – more compute always yields better models – continue indefinitely, or whether new architectures emerge that deliver more intelligence with far less compute.

AI scaling laws and what could break the boom

Today’s AI boom is anchored in empirical scaling laws: the more parameters, data, and compute you throw at a model, the better its performance. So far, bigger clusters and more tokens have reliably produced higher “IQ” models.

But these are not laws of physics – they’re empirical relationships that could change. Several things could break the current pattern:

• A fundamentally new AI architecture that achieves similar or better performance with far less compute.
• Major breakthroughs in algorithmic efficiency that slash token and compute requirements.
• Shifts in where inference runs – for example, moving more workloads to edge devices like phones and PCs.

We’ve already seen how rumors of such breakthroughs – including from China – can rattle markets, even when they don’t fully pan out. A genuine step-change in efficiency would dramatically reshape the economics of AI infrastructure and could hit today’s biggest winners hardest.

If you’re interested in how AI economics can shift away from pure software and models, it’s worth also looking at how services and implementation are emerging as major profit centers in AI projects, as explored in this breakdown of why services can be more profitable than AI software itself.

Memory: from commodity cycles to AI-driven scarcity

Memory – DRAM and NAND flash – has historically been a brutal commodity business. Producers overbuild, prices crash, and margins disappear. AI is temporarily rewriting that script.

Over the past year, DRAM and NAND prices have risen 4–5x, driven almost entirely by AI data centers. Several trends are behind the surge:

• Models have shifted from simple chatbots to more complex reasoning systems that need to store and process many more tokens.
• Context windows have expanded, allowing models to handle much larger inputs (like long documents or large codebases).
• AI agents – systems that chain multiple model calls and tools together – require even more memory and storage.

Memory makers were not prepared for this spike. They had just come through a downturn, were cautious about adding capacity, and faced physical constraints:

• Equipment makers like ASML and Applied Materials can only grow shipments about 30–35% per year due to supply chain complexity.
• New fabs and cleanroom space are expensive and take years to build.

As a result, the current upcycle in memory may be higher and longer than usual, even if it’s not “different forever.” Market valuations still imply a sharp downturn is coming soon – many memory stocks trade at mid-single-digit forward P/E multiples, as if a big price collapse is just 6–9 months away. If the cycle proves more durable, there may be room for upside.

How expensive memory is hitting PCs and smartphones

There’s a flip side to AI-driven memory demand: it’s squeezing everyone else who uses DRAM and NAND.

For PC and smartphone makers, memory used to represent roughly 20% of the bill of materials. With current prices, that share has jumped toward 50%. Many device makers operate on thin margins and can’t fully absorb that hit, so they’re raising end prices.

Consumers, however, are price-sensitive. When laptops and phones get 30–50% more expensive, many people simply delay upgrades. That’s exactly what’s happening: PC and smartphone unit volumes are down by mid-teens percentages this year, an unusually sharp decline for what are typically flat, ex-growth markets.

This creates potential short opportunities in parts of the consumer hardware supply chain – component makers that sell into PCs and phones but lack pricing power. They face falling units without the offsetting benefit of AI-driven pricing.

Semiconductor equipment: great businesses, rich valuations

Semiconductor equipment makers – the companies that supply lithography, deposition, etch, and other tools to fabs – are some of the best businesses in the ecosystem. Many are near-monopolies in their niches and have strong recurring revenue from service and consumables.

But there’s a ceiling on how fast they can grow. Physical and supply chain constraints mean they typically can’t expand shipments more than about 30% per year, even in a boom. Many of these stocks now trade at roughly mid-30s forward P/E multiples.

That’s not outrageous for 30% growers, but it does make them relatively more expensive than some of the headline AI winners. For example, some leading GPU and networking chipmakers trade at lower multiples on 2027–2028 earnings despite higher margins and stronger near-term growth.

In other words, the equipment names may be more like “Costco of semis”: high quality, but already richly valued compared to other parts of the stack.

Where investors see long and short opportunities

Putting it all together, a few themes emerge for investors trying to navigate the AI and semiconductor boom:

More attractive areas on the long side

• Leading GPU and accelerator vendors: They sit at the heart of AI training and inference, have strong pricing power, and still trade at valuations that many see as reasonable relative to growth and margins.
• High-end networking and optical/photonics: As clusters grow, moving data efficiently becomes a bottleneck. Companies that solve bandwidth and latency constraints can enjoy durable demand.
• Memory makers: Despite a huge run in share prices, many still trade at low forward multiples that assume an imminent downturn. If the AI-driven upcycle lasts longer, there may be more room to run.

Areas of skepticism and short ideas

• Neo-clouds and GPU lessors: Their business models look more like financial engineering than durable tech moats, with single-digit returns on capital even in the best of times.
• Legacy data centers and some REIT-like plays: If value continues to migrate to the chips and system design, landlords and basic colocation providers may see their economics squeezed.
• Consumer hardware supply chain: Component makers tied to PCs and smartphones, without strong pricing power, are exposed to falling units as higher memory costs push device prices up.

At the same time, some investors are wary of being outright short semiconductors as a group. Even if this ultimately proves to be a bubble, the boom could last longer than skeptics expect, and the near-term earnings power of key players is very real.

The bigger picture: an economy increasingly tied to AI

Beyond individual stocks, there’s a broader structural shift underway. The US equity market is becoming an increasingly concentrated bet on AI working – not just at the chip level, but all the way up through hyperscalers, model companies, and AI-native applications.

If the massive AI capex build-out ultimately delivers strong returns – in the form of new products, higher productivity, and durable margins – the rewards for investors could be enormous. If it doesn’t, the losses will be equally dramatic, especially for those holding the most AI-exposed indices and mega-cap names.

For builders and founders, this environment also shapes where the best opportunities lie. Rather than trying to compete directly with hyperscalers or model labs, many successful AI businesses focus on services, integration, and domain-specific solutions – a pattern explored in depth in this guide to building a profitable AI-powered app with no code.

For now, the accounting mismatch between AI infrastructure sellers and buyers, the tightness in GPUs and memory, and the uncertain durability of scaling laws all point to the same conclusion: there is real money being made, but also real risk that system-wide returns fall short of today’s expectations. Navigating that tension – with a clear eye on both the numbers and the technology – is likely to define AI investing over the next several years.