The key race: AI compute manufacturing keeps sprinting, and memory has become the sector's newest and widest bottleneck
The industry's central storyline through early August, whether physical manufacturing capacity can keep pace with AI compute demand, kept playing out at the frontier: Nvidia's Vera Rubin platform reached full production and began shipping in July, and TSMC pulled its 3nm and 2nm wafer ramps two to three months ahead of schedule. But the manufacturing gate that mattered most to the wider economy in mid-August sat one layer down the stack, in memory. Samsung, SK Hynix, and Micron, who together control more than 95 percent of global DRAM production, have spent 2026 reallocating manufacturing capacity toward the high-bandwidth memory that AI accelerators consume in bulk (producing 1GB of HBM eats roughly three to four times the wafer capacity of standard DDR5), leaving conventional DRAM and NAND for phones, laptops, and everyday electronics in the shortest supply the industry has seen this decade.
That scarcity is now a consumer-market event, not just a data-center one. IDC's latest forecast, published in early August, projects global PC shipments will fall 11.3 percent in 2026 to roughly 252.5 million units even as total market value rises to about $274 billion on the back of price increases, while smartphone shipments are expected to drop as much as 13.9 percent to their lowest volume in more than a decade, with average selling prices climbing 14 percent as the sub-$100 segment becomes largely uneconomical. Lenovo has been the most visible casualty of the squeeze: CFO Winston Cheng disclosed the company is carrying component inventories roughly 50 percent above normal levels specifically to delay passing AI-driven memory costs on to buyers, even as he called the pace of the increases unprecedented. IDC expects the shortage to persist through 2027, with prices not expected to ease meaningfully until 2028.
Who's ahead: TSMC extends its packaging and capex lead, Samsung and SK Hynix cash in on the memory supercycle, and Intel banks its first named 14A customer
TSMC kept converting demand into hard commitments: its board approved roughly $29.4 billion in fresh capital appropriations on August 11 and made its Sony image-sensor joint venture legally binding the same week, while the company began outsourcing part of its CoWoS advanced-packaging process to OSAT partners including ASE after Nvidia alone reserved an estimated 800,000 to 850,000 CoWoS wafers for 2026, more than half of TSMC's total capacity; the company's 5.5-reticle CoWoS-L package is now running at 98 to 99 percent yield in volume production. The memory makers had an even louder quarter. SK Hynix reported record second-quarter revenue of 79.3 trillion won, up 257 percent year over year, at a 76 percent operating margin, as it began mass shipment of HBM4, though the results still missed analyst expectations because some HBM4 shipments and revenue recognition slipped into the third quarter. Samsung reported HBM4 yields have reached roughly 80 percent and told investors third-quarter HBM4 revenue will more than triple sequentially, with HBM4 set to exceed 60 percent of total HBM revenue in the second half, underpinning the roughly $200 billion multiyear AI-chip supply agreement it struck with Broadcom in July spanning HBM, 2nm logic, and packaging through 2030.
Intel had its most concrete foundry news of the year. On August 10 the company priced an upsized $20 billion common-stock offering (from an initially announced $15 billion), explicitly to fund AI-driven capital spending including its unproven 14A node, and paired the raise with its first named external 14A customer: Tesla, which will use the process to manufacture AI chips for its Terafab project. Intel has committed to high-volume 14A production in 2028 and says Google, Apple, AMD, and Nvidia are evaluating commitments once its PDK 0.9 design kit becomes available later in 2026, after warning earlier in the year that the node could be shelved without a major external anchor customer.
What decides it: whether the Kyber delay holds, whether the memory shortage breaks something first, and how far China's homegrown tools actually reach
Two open questions still frame the sector's next year. The first, whether Nvidia's Kyber rack for Rubin Ultra is really delayed to 2028, as SemiAnalysis reported in early July citing a PCB midplane that cannot be manufactured at reliable yields, remains unresolved; Nvidia has said only that its roadmap is intact without confirming or denying a slip, and the ambiguity matters because Kyber was meant to be the scale-up vehicle for 2027's highest-margin AI hardware. The second, newer question is whether the memory shortage resolves on the industry's own timeline or forces a harder reckoning first. Data cited by TrendForce shows the panic-buying phase already cooling: DRAM contract prices are projected to rise 13 to 18 percent quarter over quarter in the third quarter, a sharp deceleration from roughly 60 percent gains in the second, but IDC's own forecast has the shortage running through 2027 regardless, meaning the PC and phone makers absorbing it have more than a year left to manage around it, not months.
China's parallel track kept advancing on its own terms. SMIC has effectively become the country's designated national champion for strategic AI silicon, now mass-producing Huawei's Ascend 910B accelerator and planning to double its 7nm-class capacity in 2026, while Beijing has moved to mandate that new domestic fabs source at least half their equipment locally as it targets 80 percent semiconductor self-sufficiency by 2030, up from roughly a third in 2024. The clearest proof point of that push is Shanghai Yuliangsheng Technology, a state-backed toolmaker that TrendForce reported on July 28 has begun mass-producing an immersion DUV lithography system, with first units due to SMIC, Hua Hong, and CXMT before the end of 2026 and volume rising toward roughly 20 machines in 2027. The tool is capped at 28-nanometer features in a single exposure (7-nanometer-class output is possible only through costlier, lower-yield multipatterning) and still depends on some imported Japanese components, but it is the first credible sign of a second country building the immersion lithography ASML has held alone for two decades.
The money and the rules: Nvidia turns compute into collateral, Intel and TSMC raise capital at record scale, and the MATCH Act edges toward a real vote
Capital kept arriving at a scale with no real precedent. On August 10 to 11, Nvidia partnered with Apollo Global Management, Blackstone, BlackRock, Brookfield, Goldman Sachs, and KKR to launch AI compute infrastructure financing platforms aiming to mobilize more than $500 billion of third-party capital, using compute power itself as loan collateral, a new channel layered on top of TSMC's own record $60 to $64 billion 2026 capex budget and its newly finalized 5 to 10 percent wafer-price increases across advanced nodes effective January 2027.
The rules kept tightening in parallel. The bipartisan MATCH Act, which would force US allies to align their DUV and lithography-servicing export controls with Washington's, cleared the House Foreign Affairs Committee 44-0 in April and is now positioned for inclusion in the Senate's National Defense Authorization Act, a path that would make it one of the closest things to a guaranteed floor vote a standalone export-control bill can get, since Congress has passed an NDAA every year for more than six decades. ASML's own regulatory exposure remained unresolved: as of its most recent earnings call, no formal Bureau of Industry and Security enforcement action had followed Commerce Secretary Howard Lutnick's claim that an EUV system or component may have reached China in breach of export controls, a claim ASML has categorically denied while accounting for all 314 operational and 26 decommissioned EUV units worldwide by location.
What to watch through 2027
Watch Nvidia's fiscal second-quarter earnings on August 26, guided to roughly $91 billion in revenue, for the market's next verdict on both the Kyber delay question and the AI-financing debate. Watch TSMC's mid-October Q3 report for whether 2nm capacity reaches its 100,000-wafer-per-month year-end target and whether CoWoS outsourcing to OSAT partners actually eases packaging pressure. Watch ASML's own Q3 report on October 14, guided to 11 to 12 billion euros in net sales, and watch whether Samsung, SK Hynix, or another memory maker becomes the second real High-NA EUV customer beyond Intel. Watch Intel's PDK 0.9 design kit ship later in 2026 and watch whether Google, Apple, AMD, or Nvidia convert 14A evaluation into a second named customer alongside Tesla.
Watch whether DRAM contract prices actually decelerate to the 13-18 percent quarterly pace TrendForce is projecting for the third quarter, the first real test of whether the memory shortage is cresting or merely pausing before IDC's own forecast has it running through all of 2027. Watch the MATCH Act's fate inside the Senate's NDAA process and whether the Bureau of Industry and Security brings any enforcement action against ASML. And watch China's parallel build-out: whether Shanghai Yuliangsheng's immersion DUV tool actually reaches SMIC, Hua Hong, and CXMT production lines before year-end, and whether SMIC's planned doubling of 7nm-class capacity holds, the two clearest tests of how far China's chipmaking independence push has actually gotten by the time 2027 begins.
Sources
- Tom's Hardware - IDC slashes 2026 PC shipment forecast amid memory shortages
- SK hynix - 2Q26 financial results
- TrendForce - China reportedly starts mass-producing immersion DUV tools
- SiliconANGLE - Intel launches upsized stock sale amid AI boom, advanced packaging demand
- Crypto Briefing - MATCH Act poised for inclusion in Senate NDAA
- Nvidia - Nvidia partners with Wall Street firms to mobilize over $500 billion of AI compute financing
