The strongest fact: a memory maker is spending at a new scale
The clearest verified evidence of an AI-driven memory squeeze is in Micron Technology's filings, not in forecasts. Micron's capital expenditure, meaning the money spent on factories and equipment, was $7.676 billion in fiscal 2023 and $8.386 billion in fiscal 2024.1 In fiscal 2025 it reached $15.857 billion.1 That is almost twice the prior year's figure, in a company whose spending had barely moved the year before.
The quarterly figures show the pace is still rising. Micron spent $1.796 billion in the first quarter of fiscal 2024, $3.206 billion in the same quarter of 2025, and $5.389 billion in the first quarter of 2026.1 That is about three times the 2024 level in two years. All of these figures come from SEC filings and are checked against that source.
Micron's memory chips are widely used in data centres, and the dossier's narrative links the shortage to AI demand and expects it to last 12 to 18 months. The expected duration is our narrative's framing, not a verified figure. The filings establish only that the company is spending far more.
Cash is piling up alongside the spending
Micron's cash balance was $8.577 billion at the end of fiscal 2023 and $7.041 billion at the end of fiscal 2024.2 It was $9.642 billion at the end of fiscal 2025.2 The recent quarters show a sharp climb: $13.908 billion in the second quarter of fiscal 2026 and $24.995 billion in the third quarter.2 That is roughly two and a half times the $10.163 billion held in the third quarter of fiscal 2025.2
This matters because Micron is spending much more on factories and still ending up with far more cash. Our records also note that, under CHIPS Act restrictions, Micron could technically restart stock buybacks. Analysts are not necessarily expecting an announcement. We have no filing showing a buyback plan, so treat it as a possibility only.
The filings also show costs rising. Micron's cost of revenue, the direct cost of producing what it sells, was $19.498 billion in fiscal 2024 and $22.505 billion in fiscal 2025.3 In the second quarter it rose from $5.090 billion in fiscal 2025 to $6.105 billion in fiscal 2026.3 The dossier contains no verified revenue figures, so we cannot say whether profits have widened or narrowed. Rising costs alone do not show that the shortage is making Micron richer.
Securing supply: alliances and custom chips
The other development is companies locking in access to memory. Nvidia's expansion with SK Hynix, a Korean memory maker, was reported in connection with a San Francisco summit on July 24 where roughly $950 billion in new AI agreements were reportedly signed.4 Raj Mirpuri said of the Nvidia deal: "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."4 HBM is a stacked, high-speed memory that sits next to AI processors. The $950 billion is a headline total of announced agreements, not money confirmed as spent, and the source did not break it down.
Smaller players are moving toward specialised silicon. On October 8, 2026, Zeta Global said it had begun work with partners on a specialised chip, called AIM, for what it described as a capital-light approach to delivering faster intelligence more efficiently.5 The same announcement said its Athena Inference Model was developed with Fireworks using NVIDIA Nemotron.5 Zeta is a marketing-technology company, not a chip maker, so this shows how widely the custom-silicon idea is spreading rather than a challenge to the big names.
The bottleneck is also about moving data
Memory is only half the problem. Chips must also exchange data quickly. On July 28, 2026, Bull and Kalray announced a collaboration on networking technology for AI and high-performance computing, including work on Ultra Ethernet, a new open standard.6 Kalray's Éric Baissus said the agreement confirmed "la pertinence de notre vision technologique" (the relevance of its technology vision).6 Our narrative calls the Kalray–Bull tie-up part of a consolidation trend. The excerpt we hold describes a collaboration, not a takeover, so that label should be read loosely.
On July 29, GlobalFoundries signed a letter of intent with the U.S. Department of Commerce for a $300 million award to speed up U.S. silicon photonics, the optical technology that moves data as light.7 A letter of intent is not a finished award. AMD's Mark Papermaster said: "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."7
A more radical idea comes from Hannah Earley, cofounder and chief technology officer of Vaire Computing. She argues that waste heat is a design choice, and her company is building chips that recycle energy usually lost as heat, a strategy called reversible computing.8 She describes the key component as "really a glorified pendulum."8 She also said: "I want to tackle every part of how computers are built, and rethink it in these terms."8 This is an early-stage idea from a startup founder. Notably, our measured check of this source's claims found that none of 11 held up, so we present Earley's account as her stated view, not as demonstrated performance.8
How much to trust the bullish calls
The dossier cites Wall Street optimism about memory stocks. Lynx Equity Strategies rated Sandisk a clear buy with anticipated upside of 49%, and Micron a clear buy with anticipated upside of 33%.9 These are one firm's opinions, relayed in a Motley Fool article about ASML's stock rising. Our measured reliability check found only 22% of 18 checked claims from that source held up, a small sample, but a weak result.9 Read the targets as one analyst's view, not as a forecast we have verified.
Not all consolidation reported in the sector is about AI memory. Skyworks and Qorvo, two radio-chip companies, announced the expected leadership of their combined company, pending completion of their transaction.10 Bob Bruggeworth said the announcement "reflects the strong partnership that has shaped our integration planning."10 Small firms are also benefiting from a strong sector. Solitron Devices reported net sales up 101% to about $5.44 million in its fiscal 2027 first quarter, though net bookings fell 48% to $1.45 million.11 Backlog rose 28% to $23.34 million.11 That is small in absolute terms, and it shows demand running ahead of new orders at the same moment.
Insiders are not uniformly buying. Lattice Semiconductor's SVP of Strategy and Marketing, Esam Elashmawi, reported selling 16,773 shares on Aug. 16 and 17, 2026, at a weighted average price of $134.56 per share.12 One executive's sale says little by itself, and the dossier gives no reason for it.
The friction we cannot verify
Our narrative also flags regulatory and geopolitical pressure: FTC scrutiny of AI agents, OpenAI's copying accusations against Moonshot, Chinese espionage targeting AI experts, and China's YMTC expanding NAND memory capacity. The dossier holds no sourced detail on these items, and we have not independently verified them, so we do not draw conclusions about them. The same applies to onsemi–Synaptics and DeepSeek–Huawei: they are part of the narrative but lack verified supporting material here.
What to watch
- Whether Micron's next filings show capital spending staying above the $15.857 billion fiscal 2025 level.1
- Whether its cash, now $24.995 billion, goes into buybacks or more factories.2
- Whether Micron's cost of revenue keeps rising faster than any evidence of higher sales, which would argue against a simple windfall story.3
- Whether GlobalFoundries' letter of intent becomes a finalised award.7
- Whether Zeta's AIM chip moves beyond development.5
