AI SHORTAGE TRACKER

On Thursday, July 9, 2026, CEO Gary Dickerson told Nikkei Asia in Singapore that chipmakers are sharing equipment demand outlooks for two years or more so that capacity expansion proceeds smoothly, and that Applied Materials has tremendous visibility over the next eight quarters (about two years). Some customer outlooks reach out to 2030. Supporting numbers: advanced packaging equipment revenue is expected to grow 50% this year, and the Singapore cleanroom facility will more than double. The remarks came right after the July 2 selloff in equipment stocks driven by NAND oversupply and capex-deferral fears (KLA -11.6%, Lam Research -10.2%, Teradyne -13.7%), reading as a rebuttal; on publication day Applied Materials rose 6-8%, TD Cowen raised its target to $700 and Mizuho to $650. This differs in character from the May 14 earnings-call comment about supply concerns lasting to 2030: May was about customers' supply worries, whereas this is about order visibility and long-term commitment.

On Wednesday, June 25, 2026 — hours after Micron posted a blockbuster FQ3 2026 (revenue +346%, GAAP gross margin ~84.6%) — CBO Sumit Sadana told the Wall Street Journal that some customers (without naming Apple) pushed for rock-bottom prices during the 2022-2023 downturn, turning Micron's margins negative and forcing the industry to shut down the capacity investments that could have prevented today's memory shortage. His remarks landed just after Apple rolled out sweeping hardware price hikes blamed on memory costs, fueling the blame debate. Cross-cited by 9to5Mac (6/25) and MacRumors (6/26) from the WSJ original.

At the June 18, 2026 FERC Open Meeting, Chair Swett presided over a unanimous vote ordering six major US grid operators to fast-track AI data center interconnection requests within 90 days (down from multiple years). Data centers must bear their own interconnection costs. The ruling is a formal acknowledgment that power-grid bottlenecks are the binding constraint on AI infrastructure expansion. Reported June 18 by Bloomberg, TechCrunch, and Spokesman-Review.

June 17 Bloomberg TV interview at VivaTech in Paris. Fouquet said demand for AI infrastructure is still 'enormous' and noted that large new AI fab projects such as Elon Musk's Terafab are attractive 'as long as you're not supply limited' — implying the real constraint on the AI buildout is EUV lithography supply, not demand, and that ASML must manage this carefully. Distinct from the existing ASML entry (2026 Q1: chip demand outpacing supply, customers accelerating capacity), this remark directly flags equipment supply limits on new projects. China's Cailianpress (STAR Market Daily) summarized it on 6/17 as 'AI infrastructure demand keeps climbing.'

AP exclusive interview on June 16 in Sherman, Texas, ahead of a groundbreaking for an expansion of Coherent's factory. Huang argued the real constraint on AI infrastructure is power, not chips, blasting the U.S. for suppressing energy production for too long. The expansion builds a line for lasers (optical interconnect) that transmit data among chips and could cut AI systems' power use by up to 50%. Distinct from the existing Nvidia remarks (Computex 'Please Make More', Vera Rubin, FY2027 Q1), which concern GPU/HBM supply; this is a power-bottleneck statement. The other power-grid entry (GE Vernova, gas turbines) is a separate remark.

Oracle Q4 FY2026 — RPO (remaining performance obligations) grew $85B in a single quarter to $638B. Cloud infrastructure (IaaS) revenue was $5.8B (+93% YoY). Co-CEO Magouyrk said global GPU utilization is 97.5%, and GPUs returned at contract renewal were instantly resold to other customers within the same quarter. Oracle delivered more than 1.2GW of capacity to customers this quarter — nearly matching the previous four quarters combined. Distinct from the existing Oracle Q1 FY2026 remark (Catz, $455B RPO).

Remarks to reporters after the June 4 annual general meeting in Hsinchu. Wei called this year's customer demand growth 'insane' and said the explosive growth of the AI industry has left the entire supply chain and ecosystem unprepared, with constraints everywhere — logic, memory, packaging, testing, cooling and power. He admitted TSMC itself faces equipment shortages due to component shortages at its tool suppliers, yet said the company will avoid sharp price hikes. Distinct from the existing TSMC Q1 2026 earnings remark (CoWoS packaging).

Bookings ran ~3x shipments — Broadcom's Q2 FY2026 AI semiconductor revenue was $10.8B (+143% YoY) but bookings topped $30B. Q3 AI revenue is guided to $16B (+200%+ YoY), $56B for FY2026, and over $100B for FY2027. Hock Tan said demand visibility now runs all the way to 2028, with customers placing orders early and in huge volume. Distinct from the existing Broadcom Q4 FY2025 remark.

Press briefing at Computex 2026 in Taipei (June 2). Chairman Chey reaffirmed his forecast that the AI-driven memory shortage will persist until 2030, and said SK hynix will double its total wafer capacity within five years. With greenfield fab lead times of more than five years, the new output arrives only near the tail end of the shortage window he is predicting. He spoke in Korean; the English quote is the media translation. Distinct from the existing SK hynix Q1 2026 earnings remark (HBM sold out for 3 years).

On June 2, Jensen Huang made a surprise visit to SK hynix's Computex 2026 booth and signed the displayed seventh-generation HBM (HBM4E) wafer with 'Please Make More.' He also wrote 'LOVE SOCAMM' on a 192GB SoCAMM module. It is a direct signal from the largest customer that HBM supply is short. Vera Rubin carries HBM4 from all three suppliers — SK hynix, Samsung and Micron — and is now in full production. Distinct from the existing NVIDIA FY2027 Q1 earnings remark (Vera Rubin supply constraint).

The bottleneck of the AI era is compute supply, not demand. Friar said: “If you want to buy more compute in 2026, good luck — I don't know where you'd get it,” expecting scarcity through 2026 and a still-limited 2027. The bottleneck is not a single GPU but everything — power, land, power infrastructure, regulation, racks, chip supply chain, memory, talent, and community trust — with a memory-demand surge already underway. A 1GW data center costs ~$50B fully loaded and takes ~3 years, so even the Michigan site won't deliver until late 2027/early 2028. Per-gigawatt compute cost is rising as power and memory get more expensive. A $122B raise closed in March 2026 secured ‘maximum optionality’; compute grew 0.2GW (2023)→1.9GW (2025) while OpenAI diversified beyond NVIDIA/Azure to AMD, Cerebras, a custom Broadcom design, and Oracle/CoreWeave/GCP/AWS. Where she feels shortest on compute now is 2030–2032.

Dell posted $16.1B of AI server revenue in Q1 FY2027 (+757% YoY) and exited the quarter with a $51.3B AI backlog. COO Jeff Clarke stressed the bottleneck is memory (DRAM and NAND), not GPUs. The company is supply-constrained in the second half and framed it explicitly as a supply, not demand, problem — defending margins through cost control amid surging memory prices.

Marvell Q1 FY2027 revenue was a record $2.42B (+28% YoY). It raised its data center interconnect growth outlook from 50% to more than 70% year over year. Murphy said demand keeps accelerating and 'every sort of program we looked at a year ago is larger when we look a year later,' with supply-securing investments enabling the business to scale every quarter.

Demand for the next-gen Vera Rubin platform outpaces capacity from the start of shipments. Said in response to a question from Vivek Arya.

Demand-side signal — monthly token volume jumped 7x, from 9.7T two years ago to 480T last year to over 3.2 quadrillion now. Many companies have blown through their annual token budgets, and it's only May.

Largely sold out of 2026 capacity, with prices rising across Ampere, Hopper and Blackwell. Revenue backlog $99.4B (~4x YoY).

Arista Networks Q1 2026 earnings call, May 5, 2026. Ullal added: 'Clearly, our demand is outstripping our supply this year.' and 'At first, we thought it was memory. Now it's all the wafer fabrication facilities. Every chip is challenged. I think the supply chain problem is not a one- or two-quarter phenomena. We now think it's a one- or two-year phenomena.' Arista is the dominant supplier of AI data-center networking (Ethernet switches/routers) with 80%+ share in hyperscaler direct-connect networks. No prior Arista entry in tracker.

MI450 / Helios rack-scale deployment demand now exceeds initial plans. Set against the Meta 6GW Instinct deployment and the OpenAI partnership.

AI demand deepens the memory shortfall as HBM4 enters mass production; fulfillment rate at a record low. The English quote is a media rendering of the Korean-language call.

Shifts to a new business model backed by multi-year (up to 5-year) customer commitments. Five partnerships, $42B committed, with over a third of FY2027 bits pre-committed; revenue +97% QoQ.

Meta CFO Susan Li acknowledged the company has kept underestimating its compute needs even while ramping capacity aggressively. It raised 2026 capex guidance to $125-145B, reflecting 'expectations for higher component pricing this year and, to a lesser extent, additional data center costs to support future-year capacity' — a sign that surging memory prices are directly lifting hyperscaler capex.

On the Q1 2026 call, CEO Andy Jassy said component costs — memory in particular — had skyrocketed and that there simply isn't enough capacity for the demand. AWS AI revenue topped $15B in the quarter but could have grown faster absent capacity constraints. In-house Trainium2 is essentially sold out and Trainium3 nearly fully subscribed. AWS added 3.9GW of power in 2025 and plans to double total power capacity by end of 2027.

Three years of HBM demand exceeds supply — effectively three years sold out. The English quote is a media rendering of the Korean-language call.

Gas turbine lead times of about three years, with 2030 slots already selling and backlog up $13B in the quarter. (Original context: turbines are 'not the gating item' — 'supply constrained' is not a verbatim phrasing.)

Wei acknowledged advanced packaging (CoWoS) capacity is very tight. Said in response to a question about larger reticle sizes and competitive threats.

Chip demand outpaces supply. Year-end 2025 backlog EUR 38.8B; Q4 bookings EUR 13.2B (~2x consensus). Memory systems revenue topped logic for the first time.

Both DRAM and NAND demand exceed supply. Management said it can fulfill only 50% to two-thirds of some key customers' demand in the medium term, and expects tightness to persist beyond calendar 2026.

Leading-edge logic and DRAM capacity essentially full with prices rising. AI HBM DRAM needs 3-4x more wafers per bit than standard DRAM. DRAM rose from 27% to 34% of Semiconductor Systems revenue.

Demand/architecture-shift remark — legacy infrastructure isn't built for AI and needs a rethink. A demand narrative rather than a strict shortage quote. (Patel's 'chatbot to agent' framing was at Cisco Live, not the call, so it is excluded.)

Data center book-to-bill above 4x, with most of 2026 and much of 2027 already booked. Strong 800G and 1.6T transceiver demand; the imbalance is expected to persist next year. Distinct from the existing FY26 Q1 (InP laser) remark.

CFO formally states it could not fulfill demand — effectively sold out — and expects the next quarter to be even more undersupplied than Q2.

Entire 2026 HBM volume and price pre-committed. Said it can meet only 50-67% of some key customers' demand in the medium term (across memory broadly). Distinct from the existing FY26 Q2 NAND remark.

Demand-side signal — AI semiconductor revenue surged 74%, guided to double to $8.2B next quarter, driven by custom AI accelerators and Ethernet switches.

Extends the undersupply outlook beyond 2027; allocation escalates from several products to all end markets, with customers pre-committing 2027 demand.

Indium phosphide (InP) laser capacity — particularly EMLs — for optical transceivers capped data center revenue growth. Coherent is ramping a 6-inch line to expand internal InP capacity.

Azure demand far exceeds available capacity; expects to stay capacity constrained through the end of FY2026. Same diagnosis two quarters running.

OCI consumption revenue +57% with demand far exceeding supply. RPO backlog up 359% to $455 billion.

First official 'demand exceeded supply' since the spin-off; allocation begins on several products at the same time.

First independent earnings call after the spin-off from Western Digital. A turning point as it deliberately cut supply and began price increases, exiting NAND oversupply.