On July 24, 2026, Jensen Huang, Sam Altman and the heads of Samsung, SK Group, Hyundai Motor and Naver converged on San Francisco for a summit with South Korea's president — and by the end of the day, roughly $950 billion in new AI agreements had been signed.1 The AI chip buildout, in other words, has not paused. The stock market pricing it has: the SOXX semiconductor index is down 25% from its highs, and leveraged semiconductor ETFs have lost 62% of their value over the same span.2 For an audience tracking what AI infrastructure is actually being built, that gap — deal flow accelerating while valuations collapse — is the story.
The Buildout Keeps Signing Deals
Nvidia used the Korea summit to expand its supply relationship with SK Hynix. Executive Raj Mirpuri said the expansion "will include a co-develop opportunity for us on the next-generation SK Hynix AI memory, and this will help us secure a stable supply of HBM memory."1 That sentence captures Nvidia's current posture: it isn't just buying finished chips, it's locking down the high-bandwidth memory supply chain years out, because HBM capacity — not GPU compute alone — has become the binding constraint on how many AI accelerators can actually ship.
Nvidia's rival is moving on a parallel track. AMD and Cerebras announced a collaboration combining AMD's Instinct GPUs with the Cerebras Wafer-Scale Engine into a single disaggregated inference workflow; Cerebras plans to deploy AMD's Helios systems in its own data centers, with the joint offering available first through Cerebras Cloud in the second half of 2026.3 Where Nvidia is buying up memory supply, AMD is buying speed-to-market on inference — the deployment stage that increasingly determines enterprise wins, not just training-chip benchmarks.
The interconnect layer is drawing its own government-backed commitments. GlobalFoundries signed a letter of intent with the U.S. Department of Commerce for a $300 million CHIPS award targeting "next-generation silicon photonics — the optical technology that moves data at the speed of light and underpins the AI and high-performance computing data centers driving the global economy."4 AMD's Mark Papermaster, welcoming the deal, put the stakes in engineering terms: "As AI systems scale, moving data efficiently is as critical as increasing compute performance. Silicon photonics and advanced packaging will be key to delivering the bandwidth, energy efficiency, and system-level connectivity required for the next generation of AI cluster infrastructure."4 France's Kalray struck a parallel networking partnership with Bull the same week; CEO Éric Baissus called it validation of the company's direction: "Cet accord avec Bull confirme une nouvelle fois la pertinence de notre vision technologique et la qualité des innovations développées par nos équipes" — this agreement with Bull confirms once again the relevance of our technological vision and the quality of our teams' innovations.5
The buildout is reaching into consumer hardware, too. Phison and Intel announced a collaboration to extend the memory available to AI workloads on Intel's AI PC platforms using Phison's aiDAPTIV technology. Phison's KS Pua described the shift underway: "AI PCs are evolving into platforms for more sophisticated local AI workloads, including agentic applications and larger MoE models that place increasing demands on memory capacity and responsiveness."6 That's a notable signal in its own right — the industry now expects meaningful AI inference to run locally on PCs, not only in the cloud, which is a different capacity-planning problem than the data-center buildout usually gets credit for.
Nvidia's own ecosystem shows how far this buildout now reaches beyond chips themselves. Via News's verified entity graph traces Nvidia platforms — the NeMo Agent Toolkit, NeMo Guardrails, NeMo microservices, the Nemotron 3 Super model and DGX Spark hardware — into deployments at customers including Mount Sinai Health System, Yum! Brands and the State of Alaska Legislative Affairs Agency, with distributor TD Synnex supplying Nvidia hardware downstream and AMD listed as its direct competitor.7 That breadth — a health system, a state government agency, a fast-food chain — is part of why chip demand forecasts have stayed resilient even as chip stocks have not.
Consolidation Moves Forward
Parallel to the deal-making, the sector is consolidating. Skyworks Solutions and Qorvo announced the expected executive leadership team for their combined company, effective once the merger closes. Qorvo's Bob Bruggeworth said the announcement "reflects the strong partnership that has shaped our integration planning efforts from the very beginning," adding that he is "confident these leaders will help foster collaboration across our teams..."8 Skyworks and Qorvo both supply RF and connectivity components rather than AI accelerators directly, but the merger is a bet that scale matters as AI-driven demand ripples through the broader chip supply chain.
Governance is being reshaped around the same narrative elsewhere. Allegro MicroSystems appointed Brian White, a veteran semiconductor-industry CFO, to its board; chair Joseph Martin said White's "public-company CFO perspective, semiconductor industry experience and governance background make him a strong addition to our Board."9 Allegro describes itself as leveraging "more than three decades of expertise in magnetic sensing and power ICs to propel electrification, automation, AI data center, and robotics forward with solutions that enhance efficiency, performance and sustainability."9 Allegro is a smaller company than Nvidia or AMD, but naming AI data centers explicitly alongside automation and robotics shows how far the AI label has spread across chip subsectors that make nothing resembling a GPU.
Worth flagging plainly: these announcements were carried as company press releases on the NewsEOD wire, whose claims have checked out at rates as low as 23% and as high as 48% in Via News's own fidelity audits of that source. That doesn't mean the underlying events — a signed LOI, a completed board appointment — are false; it means a company's framing of its own merger's promise or a partnership's ambitions is the company's framing, not an independently audited outcome.
The Correction, in Numbers
Against that backdrop of continued deal-making, equities have moved sharply the other way. The SOXX semiconductor index is down 25% from its highs and leveraged semiconductor ETFs have fallen 62% from their peaks2 — steep enough to erase most of the sector's 2026 gains within weeks. The swing shows up starkly in ProShares' 2x-leveraged Ultra Semiconductors fund (USD): as of June 5, 2026, it "is up about 69% year to date, from $52.45 on December 31, and 196% over one year, from $29.94 on June 5, 2025" — before a separate report just two days later warned the same fund "implodes 17%," flagging that conditions could worsen further.10 That reversal, inside a single week, is the leveraged-ETF mechanics of the broader correction playing out in miniature: a year of gains erased at multiples of the underlying index's move.
Individual earnings have added to the pressure. On July 28, 2026, the Nasdaq 100 fell to a 2.75-month low "amid a deepening rout in chipmakers and AI-infrastructure stocks."11 The same day, Amkor Technology forecast third-quarter net sales of $1.95 billion to $2.05 billion — weaker than the $2.11 billion Wall Street had expected.11 Micron Technology's stock told a similar story at larger scale: after falling as much as nearly 40% from its all-time high starting in July 2026, it bounced 10% off its lows by August 9, though it remained roughly 30% below that peak.12
Not every name in the space is under equal pressure. Vicor Corporation, whose power-delivery components feed AI data-center customers, reported trailing-twelve-month revenue of $471.7 million and net income of $136.7 million, with a market capitalization of $11.3 billion as of the July 6, 2026 close.13 Its lead computing customer is due to transition from Gen 4 to Gen 5/VPD technology in the second half of 2026, with a ramp expected before year end.13 Small-cap defense-and-aerospace chipmaker Solitron Devices posted net sales up 101% to roughly $5.44 million in its fiscal 2027 first quarter, with backlog up 28% to $23.34 million, and said it expects sales "to continue to be at this level or greater for the remainder of the 2027 fiscal year."14 Its net income, notably, was dented by a small accounting item: "Net income in the fiscal 2027 first quarter was impacted by an increase to the contingent consideration of $0.33 million related to the MEI earn out payment based on a large order received after the end of the fiscal quarter."14 It's a reminder that even a growth story carries friction most AI headlines skip past.
Reading the Divergence
Via News's analysis of this data suggests the split between deal flow and share prices reads as a late-cycle correction rather than a reversal of the AI buildout itself: every major disclosed transaction here — Nvidia-SK Hynix, AMD-Cerebras, GlobalFoundries-DoC, Bull-Kalray, Phison-Intel — points toward more AI infrastructure capacity being committed, not less, with delivery timelines running through 2026 and 2027. What appears to have changed is the market's willingness to pay up in advance of that capacity actually shipping. Micron's bounce off its lows is the most concrete data point suggesting sentiment may be stabilizing, though at 30% below its all-time high it is far from a full recovery, and it is one company's stock rather than a sector-wide signal.12 Amkor's guidance cut, landing the same week as the broader Nasdaq rout, is the more cautionary data point: real revenue expectations, not just sentiment, are being trimmed at some suppliers even as others sign fresh partnerships.11
What to Watch
A few dated milestones will test which read holds up. Vicor's lead customer is expected to ramp Gen 5/VPD technology before the end of 202613 — a concrete delivery point rather than a guidance promise. The AMD-Cerebras inference offering is due through Cerebras Cloud in the second half of 2026.3 The Skyworks-Qorvo leadership team takes effect only once that merger formally closes.8 And Amkor's next quarterly numbers will show whether its guidance cut was an isolated miss or the start of a broader downgrade cycle across the assembly-and-test layer of the chip supply chain.11 None of these resolve the tension on their own, but together they're the concrete markers that will show whether the deal-making documented here converts into shipped, revenue-generating AI infrastructure — or whether the equity market's skepticism turns out to have been early.