The semiconductor selloff of July 2026 wasn’t a correction — it was a rotation. Capital is flowing from training infrastructure to inference, data pipelines, and open-source models. The AI investment cycle has entered its second phase, and the winners look different now.
In the first two weeks of July 2026, something shifted beneath the surface of the AI trade. On the face of it, it looked like a standard correction: chip stocks pulled back, leveraged ETFs got crushed, and headlines asked whether the AI rally had run out of steam.
The data tells a more specific story:
| Metric | Peak (June 2026) | July 2026 | Drawdown |
|---|---|---|---|
| Philadelphia SE Semiconductor Index | All-time high | Down from peak | ~20% |
| Direxion Daily Semiconductor Bull 3X ETF | All-time high | Down from peak | >50% |
| Korean leveraged chip ETF (SAMSUNG KODEX) | All-time high | Down from peak | >60% |
| SK Hynix (Nasdaq debut) | Offering price | Single-day drop | 15.4% |
| S&P 500 Momentum Index | Outperforming S&P 500 2:1 | July pullback | 11% |
The momentum names — the stocks that had powered portfolio returns through much of 2025 and early 2026 — were the hardest hit. The broader S&P 500 fell less than 1% in the same period. This was not a market-wide selloff. It was a targeted rotation out of AI infrastructure.
But here’s what the headlines missed: the rotation wasn’t out of AI. It was within AI — from one phase of the investment cycle to the next.
The first phase of the AI investment cycle was unambiguous: buy GPUs, build data centres, train models. Nvidia became the first company to surpass US$5 trillion in market capitalisation in October 2025.[11] The thesis was simple — whoever controls the compute controls the future. Chipmakers, hyperscalers, and data centre REITs all rode the wave.
That phase produced exceptional returns. The semiconductor sector was up more than 60% year-to-date even as July’s selloff unfolded. But investment cycles don’t stay in one phase forever. As training infrastructure scaled, two things happened simultaneously:
Diminishing returns on training. The gap between frontier models narrowed. Each new generation of model required exponentially more compute for incrementally better performance. The cost curve for training was becoming unsustainable.
Inference demand outpaced training demand. Once models are trained, they need to run — repeatedly, at scale, for millions of users. Inference is where the recurring revenue lives. Training is a capital expense; inference is an operating expense. The market began to price that distinction.
Three data points from this week crystallise the shift:
$400 million for inference chips. General Compute, an AI inference cloud startup, landed a $400 million loan from Upper90 — reportedly the first deal to put up inference-specific chips as collateral.[1] The chips aren’t Nvidia GPUs. They’re SambaNova SN50 chips designed specifically for inference: power-efficient, no water-cooling required, 16x faster inference than GPU-based clouds.[1] Upper90’s CEO Billy Libby put it bluntly: “Everyone doesn’t need a supercomputer, but they do need inference and AI.”[1]
Databricks at $188 billion.[2] The data platform company hit a $188 billion valuation in a round led by Coatue — up from $134 billion just five months prior.[2][3] Databricks isn’t a chipmaker. It’s a data-to-AI pipeline company. Its thesis: enterprises already have their data on the platform, and the next AI play is connecting that data to models efficiently. CEO Ali Ghodsi publicly benchmarked internal AI costs for his 3,000 engineers and championed affordable open-weight models (specifically Chinese GLM 5.2) for coding tasks. The message: cost control matters more than model capability.
Source: Databricks — “Databricks is Raising a Strategic Round of Funding at a $188 Billion Valuation” (company announcement, 2026). databricks.com
Apple overtakes Nvidia. On July 17, Apple reclaimed the title of world’s most valuable company (US$4.88T vs Nvidia’s US$4.86T).[4] The market is rewarding platform integrators that monetise AI end-to-end over pure infrastructure suppliers. As one analyst put it: “The re-rating reflects confidence in earnings durability rather than speculative AI upside.”
| Company | Play | July 2026 Signal |
|---|---|---|
| Nvidia | Training GPUs | Overtaken by Apple at #1 |
| Databricks | Data pipelines | $188B valuation (+40% in 5 months) |
| General Compute | Inference neocloud | $400M financing closed |
| Apple | AI platform integration | World’s most valuable company again |
| OpenRouter / Fireworks | Open model access | Raising at huge valuations |
The inference pivot is inseparable from the open-source movement. As training costs became astronomical, enterprises needed a way to run AI without paying frontier model prices per token. The answer: open-weight models that can be deployed on cheaper inference hardware.
Databricks is the poster child. The company published research on cost savings from open-weight models for coding, specifically championing Z.ai’s GLM 5.2 — a Chinese model.[10] The pattern is clear: use open-weight models for high-volume, cost-sensitive workloads. Reserve frontier models (GPT-5.6, Claude Fable 5) for tasks where quality justifies the premium.
Chinese AI labs are accelerating this trend. Moonshot released Kimi K3 on July 17, claiming benchmarks comparable to US labs’ best offerings — outperforming all rivals except Anthropic’s Claude Fable 5[9] and OpenAI’s GPT-5.6.[7][8] The model is priced at roughly Anthropic Sonnet levels, allowing Moonshot to charge a premium over other Chinese models. The implication: the US moat in AI capability is narrowing faster than most investors price in.
George Yeo, Singapore’s former foreign minister, put it directly: the open-source shift could hit US AI valuations. When capability converges and cost diverges, the premium for closed models compresses.
For investors with AI exposure, the rotation creates both risk and opportunity. Three implications:
Semiconductor concentration risk is real. The leveraged ETF carnage in Korea — the SAMSUNG KODEX SK Hynix ETF fell 45% from its debut price, over 60% from its June peak — is a warning about leverage amplifying drawdowns in a concentrated sector.[6] Goldman Sachs flagged built-up leverage across the marketplace: rising retail margin, growing levered ETF AUM, surging short-dated options volume. If you have AI exposure through levered instruments, the math is against you in a rotation.
The AI trade is broadening, not closing. Capital is rotating from pure chip plays into data infrastructure, inference platforms, and application-layer companies. Databricks, Apple, OpenRouter — these are all AI beneficiaries, just further down the value chain. For portfolio construction, this means the AI allocation isn’t going away; it’s diversifying.
Geographic diversification within AI matters more. Chinese AI models closing the capability gap introduces a new variable: if open-weight models from China are “good enough” for enterprise workloads, the total addressable market for US frontier models shrinks. This doesn’t mean US AI companies fail — it means their pricing power faces structural headwinds. For Singapore-based investors, it means the AI allocation can’t be a blind bet on US tech.
Singapore is a power- and land-constrained data centre jurisdiction, which makes it structurally an inference market rather than a training one. The Green Data Centre Roadmap, built on the Digital Connectivity Blueprint launched in June 2023, targets at least 300 megawatts of additional data centre capacity in the near term, with more to follow through green energy deployments[12]. Capacity is allocated against efficiency and sustainability criteria, not simply to the highest bidder.
Efficiency is what keeps that growth inside the power budget: the Singapore Standard on Energy Efficiency of Data Centre IT Equipment (SS 715:2025) targets at least 30% savings in the energy consumption of data centre IT equipment[12]. For a country with no frontier training cluster, the pivot from training to infrastructure for inference is not a thesis to be argued — it is the only version of AI infrastructure the island can host at scale.
Not everyone is convinced the rotation is structural. TSMC reported a 36% surge in sales in June — fresh evidence that AI infrastructure spending remains robust.[5] The semiconductor index, despite the 20% pullback, is still up more than 60% year-to-date. As one market veteran told CNA: “I don’t think it has really anything to do with fundamentals as much as just repositioning of portfolios and taking profits in stocks that have gone crazy.”
That’s a fair point. The chip selloff has a profit-taking component. But profit-taking and structural rotation aren’t mutually exclusive. Markets can reprice a sector while the underlying businesses remain healthy. In fact, that’s what a healthy market does — it arbitrages between what’s priced for perfection and what’s priced for reality.
The question isn’t whether AI spending will continue. It’s whether the next dollar of AI spending goes into training infrastructure or inference infrastructure. The evidence from July 2026 points decisively to the latter.
Three indicators that will show whether this rotation holds:
1. Inference chip adoption. If General Compute’s model — inference-specific silicon, non-Nvidia, backed by chips-as-collateral financing — becomes a template, it validates the pivot. Watch for follow-on deals from Upper90 and competitors.
2. Open-weight model market share. Databricks publicly benchmarked cost savings from Chinese open-weight models. If more enterprises follow suit, the pricing pressure on closed models accelerates. The Moonshot Kimi K3 launch is a data point in this direction.
3. Leverage unwinding. The Direxion 3X ETF and Korean leveraged chip funds are canaries in the coal mine. If they continue to lose AUM, it means retail and institutional money is genuinely rotating out of leveraged AI exposure — not just taking profits.
The AI investment story isn’t over. It’s just entering its second act — and the script is different now.
This analysis is based on publicly available data as of 2026-07-18. For related coverage, see The AI Capex Reckoning and The Capital Share.