Analysis
2026 has been a breakout year for AI hardware, with demand driven mainly by data centers. Hardware here includes semiconductors, memory chips, server racks, switches, fiber optics, silicon packaging, liquid cooling systems, and all kinds of associated energy and industrial products and services. The market’s attention has swung toward this group of assets used to build the physical infrastructure behind new AI technology.
By contrast, software demand hasn’t seen that kind of surge because it isn’t physical infrastructure. Software is code, programs, computational rules, and solutions—an industry where you don’t see supply bottlenecks.
You’ve probably noticed it too: this year, when hardware stocks rally, software stocks often fall, and vice versa. That’s a broad trend, not a strict rule; occasionally they fall together in certain stretches. Specifically:
In the first half, hardware absolutely dominated; in July and August software clearly took the baton; and in September capital started rotating back into hardware.
Looking at sector ETFs, as of September 11, 2026:
In terms of cadence:
January–June: Hardware’s Main Rally
Capital flows mainly targeted GPUs, semiconductors, servers, networking, and AI infrastructure CapEx, while software clearly lagged.
July–August: Software’s First Real Handoff
SOXX pulled back noticeably while IGV rebounded quickly. Capital began shifting from “selling picks and shovels” to “who can turn AI into revenue,” and software saw a clear valuation repair and catch-up rally.
September: Hardware Regains the Upper Hand
The software rebound began to cool, and the market tilted back toward semiconductors and the infrastructure complex.
So for now, AI hardware remains the primary trend, while AI software has entered a rotation window. A real software handoff requires AI application revenue, profits, and free cash flow to materialize on a sustained basis—not just a capital rotation after hardware has already rallied.
Hardware stocks have outperformed software stocks this year (and I believe this trend will continue). The core reason is that hardware products are more essential to AI development. Take the training of the most widely used large language models as an example: training requires thousands of GPUs running continuously for months and consumes enormous amounts of electricity, driving the entire supply chain: NVIDIA’s chip design, TSM’s wafer fabs, ASML’s lithography machines, memory from Micron and Sandisk, Celestica’s server assembly for supporting switches, and Lumentum Holdings Inc.’s lasers. These are just a small part of the supply chain.
Demand is so strong that NVIDIA’s GPU supply is scarce, which directly gives rise to direct competitors such as AMD and Intel; at the same time, indirect competitors such as cloud providers are also developing their own chips. In-house development requires custom chips, and Broadcom Inc.’s ASIC custom chips are the core of this supply chain.
Even with capacity continuing to expand, hardware remains the supply bottleneck. Building out the full hardware capacity cannot be done overnight and often takes years. Hardware scarcity has even spawned a brand-new category: new cloud service providers, represented by Nebius. These smaller companies sell compute capacity to fill the compute gap before the large cloud providers’ data centers are completed and brought online.
Data centers are another proof that hardware is indispensable: these facilities account for more than 50% of the major cloud providers’ capital expenditure. In 2026, the major cloud giants plan to spend $650 billion on data centers (compute infrastructure). This is undoubtedly a major bottleneck because it is the physical engineering required to deploy AI at scale. The latest estimates show that 24GW of new compute capacity is planned for 2026. Although a single data center can be completed within 18 months, the supporting power infrastructure often requires a 3-to-5-year construction cycle to run stably and optimally without affecting other users.
At this point, you should understand hardware’s essential role in the AI wave. On the software side, by contrast, it is hard to see hard constraints of the same magnitude. That isn’t to say software has no value — for example, when enterprises deploy AI agents, there will be demand for hardware-software integration. But at this stage, AI is more of a threat to the software industry than a growth catalyst.
A recent survey confirms this judgment: among surveyed companies, 32% stopped purchasing software products or feature modules over the past year and instead built them in-house using agentic coding technology; within the tech industry, that share rises to 41%.
Put simply, companies can now use AI to replicate code and build software in-house within weeks, whereas in the past such software could only be bought externally. But companies cannot build a data center, the supporting power, silicon lines, and GPUs in a matter of weeks—that is the most fundamental difference.
In other words: AI lowers the barriers to entry in the software industry. With just a little programming background, almost anyone can develop large projects; but AI raises the barriers to entry in the hardware industry, because all participants are competing for the same chips, power, and physical data center space.
The McKinsey survey mentioned earlier contained plenty of negative signals, but there is a potential opportunity in AI agents: among surveyed companies with annual revenue above $1 billion, 40% have already deployed AI agents at scale, up from only 27% last year.
This direction creates a transformation opportunity for the software industry: integrating AI agents into products and providing an agent runtime platform. Salesforce is the textbook case of this strategy. On the back of Agentforce and Data Cloud, it delivered a strong earnings report and sharply raised its ARR guidance by 210% to $3.9 billion; within that, the Agentforce segment’s ARR growth was even higher, at 240%. Combined with the Anthropic partnership news, the stock jumped.
For now, how much this revenue can pull through to overall revenue and profit is still unclear. Agentforce + Data Cloud together have $3.9 billion in ARR, just 9% of its full-year expected total revenue of $46 billion. There is another issue worth watching: margins. The company has already lowered its margin guidance from 20.6% to 20.1%.
Margins are critical. They directly reflect the commercialization power of AI use cases, and they are the biggest question this McKinsey report leaves for software companies. The survey shows that enterprise AI penetration has risen sharply over the past year: 88% of surveyed companies regularly use at least one AI feature. But only 37% of companies say AI has a positive impact on EBIT, flat versus last year.
One more point: token usage costs for AI agents should not be underestimated. 20% of surveyed companies say rising token costs are limiting AI product deployment. This is a new metric that did not appear in last year’s survey, so the subsequent price changes need to be tracked continuously. If software companies want to offset cost pressure, they must raise service pricing. But real-world resistance is substantial: 32% of companies have already stopped buying external software modules and shifted to building AI internally.
Hardware companies, by contrast, have stronger pricing power.
From semiconductors to memory, product prices are rising across the entire supply chain, and eventually that will be passed through to end consumers. Apple’s product price increases from a couple of days ago are a case in point. Software companies will most likely also need to raise prices in the future, because they will consume more tokens and invest more in R&D. Companies that can execute this well will claim a seat in the AI agent race.
The AI hardware sector has surged over the past year or two, and its earnings are highly dependent on a small number of customers—namely the large cloud providers and their massive capex. That is the core investment risk: if capex is suddenly cut sharply, revenue and margins will inevitably come under pressure.
But if capex declines gradually, the risk is manageable. Companies can rely on collecting payments from existing orders and optimizing cash flow; an orderly decline in capex will not trigger a crash or severely hit the AI hardware sector. The real risk is a cliff-like collapse in capex.