The AI buildout depends on memory chips as much as on processors. Memory is where models and the data they work on are held and fed to the compute. The clearest verified evidence of how hard that part of the supply chain is being pushed is in Micron Technology's SEC filings, not in analyst commentary. Micron's capital expenditure, the money spent on factories and equipment, was $7.676 billion in fiscal 2023 and $8.386 billion in fiscal 2024.1 It then rose to $15.857 billion in fiscal 2025, more than double the fiscal 2023 level.1 The quarterly figures show the same direction: $1.796 billion in the first quarter of fiscal 2024, $3.206 billion a year later, and $5.389 billion in the first quarter of fiscal 2026.1
What the filings show, and what they don't
Micron's balance sheet moved in the same period. Cash was $8.577 billion at the end of fiscal 2023, $7.041 billion at fiscal 2024 and $9.642 billion at fiscal 2025.1 By the third quarter of fiscal 2026 it was $24.995 billion, against $10.163 billion in the same quarter a year earlier.1 The second quarter of fiscal 2026 stood at $13.908 billion.1
The cost of producing what it sells also grew. Cost of revenue was $16.956 billion in fiscal 2023, $19.498 billion in fiscal 2024 and $22.505 billion in fiscal 2025.1 The quarterly figure was $5.997 billion in the first quarter of fiscal 2026 and $6.105 billion in the second.1 Our dossier holds no revenue figures for Micron, so we cannot say from these filings how profitable the surge has been. What the numbers do show is a company spending heavily and holding far more cash, which fits a market where demand is outrunning supply.
The analyst view, with a caveat
Our tracked narrative says AI-driven demand is producing a chip shortage that analysts expect to last 12 to 18 months. It is Via News's synthesis, drawn from multiple documents, rather than a single sourced forecast. The research firm Lynx Equity Strategies rated Micron a clear buy with anticipated upside of 33%, and rated Sandisk a clear buy with anticipated upside of 49%.2 That report was covered in a Motley Fool article on ASML's share price. Our checks found that only 22% of 18 claims from that outlet held up, so we treat the ratings as an analyst's opinion and not as a forecast we have verified.2 The narrative also lists SK Hynix among the memory makers lifted by the shortage, but nothing in our source documents backs a rating for it, so we do not report one.
Locking in supply
Memory is scarce enough that buyers are negotiating for it directly. In reporting on a South Korea summit in San Francisco on July 24, Nvidia's Raj Mirpuri said of an expansion with SK Hynix: "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."3 The article says roughly $950 billion in new AI agreements were signed that day.3 That is the headline claim from a source whose measured reliability is 57% of 4,956 checked claims, so it is more dependable than some, but not a source we would rest the figure on alone.3
An AI-model builder being willing to co-develop memory with a supplier, rather than just buy it, suggests the bottleneck is not only a matter of price.
Consolidation and the networking layer
Several deals show the industry reshaping around AI. Skyworks and Qorvo announced the expected leadership of their combined company on July 28. Qorvo's Bob Bruggeworth said: "Today's announcement reflects the strong partnership that has shaped our integration planning efforts from the very beginning."4 On the same day Bull and Kalray announced a collaboration on networking for AI and high-performance computing, including compatibility with Ultra Ethernet, a new open standard.5 Kalray's Éric Baissus said the agreement "confirme une nouvelle fois la pertinence de notre vision technologique".5 That source scores 30% of 2,927 checked claims, so we rely on the wording of the announcement and not on any outside claim in it.5
Data movement is a recurring theme. AMD's Mark Papermaster said: "As AI systems scale, moving data efficiently is as critical as increasing compute performance."6 He was commenting on GlobalFoundries' letter of intent with the U.S. Department of Commerce for a $300 million award to advance silicon photonics, which uses light to move data.6 Our records also show onsemi and Synaptics amending their June 25, 2026 merger agreement on October 1 after an unsolicited competing proposal from a third party. The amended deal values Synaptics at about $5.7 billion, or $123 per share in cash, and is lower than the original value.7
A different bet: chips that recycle energy
Not every effort aims to make chips faster. Hannah Earley, cofounder and CTO of Vaire Computing, is building chips around reversible computing, which recycles energy normally lost as heat.8 She says: "I want to tackle every part of how computers are built, and rethink it in these terms."8 She also describes conventional computing as "racing through a city only to pump the brakes at every intersection", where the car loses momentum and must burn more fuel to accelerate again.8 Her description of the resonator is more modest: "It's really a glorified pendulum."8
This is the weakest-sourced piece of the story. Only 0% of 11 checked claims from the MIT Technology Review document held up in our measurement, though the quotations above are attributed directly to Earley.8 The idea is early research, and the dossier carries no performance data. Our narrative also tracks Apple's possible entry into enterprise AI servers, Google TPUs, Cerebras wafer-scale chips and chip-software work at DeepSeek and Huawei, but the dossier has no source documents for them, so we make no claims about them here.
Friction and the periphery
Our records show OpenAI accusing China's Moonshot AI of copying its models on October 1.9 The narrative also flags Chinese espionage aimed at AI experts, an FTC probe of AI agents and Pentagon AI programs as sources of risk. We have no source documents for those three, so they remain unverified context.
Smaller defense-chip companies are largely peripheral. Solitron Devices reported net sales up 101% to about $5.44 million for its fiscal 2027 first quarter, against $2.70 million a year earlier. Net bookings fell 48% to $1.45 million.10 Its backlog rose 28% to $23.34 million, and it expects revenue "to continue to be at this level or greater for the remainder of the 2027 fiscal year".10 At that scale it shows AI-era demand rippling through the wider chip sector, not driving it.
What to watch
Micron's capex and cash figures in coming filings will show whether spending keeps climbing or levels off. Whether the amended onsemi and Synaptics deal closes is the next test of consolidation. The 12-to-18-month shortage estimate is the one to hold loosely, because we could not trace it to a verified source. Reversible computing needs measured results before it counts as anything more than a promising idea.