Nvidia's problem is memory, not just GPUs
Nvidia's Raj Mirpuri said the planned Nvidia-SK Hynix 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 The company that sets the pace for AI hardware is therefore getting involved in designing its supplier's memory in order to guarantee supply. The same report says Nvidia has spent much of 2026 signing deals across Asia, the Middle East and Europe to secure chips, memory and power, and that roughly $950 billion in new AI agreements were signed at a July 24 summit in San Francisco.1 The dossier gives no yardstick for that $950 billion, so treat it as a headline figure and not a measured outcome. Reliability matters here too: only 56% of 4,956 claims checked from that source held up.1
Our topic framing is that this supply tension is expected to last 12 to 18 months. That is Via News's synthesis of the material, not a figure from any single document above.
The verified numbers: Micron is spending at a different scale
The cleanest evidence is in Micron's SEC filings, which we have checked against the source. Micron's annual capital expenditure was $7.676 billion in fiscal 20232 and $8.386 billion in fiscal 2024.2 In fiscal 2025 it reached $15.857 billion.2 By our arithmetic, that is about 1.9 times the prior year.
The quarterly figures show the same direction. First-quarter capex was $1.796 billion in the first quarter of fiscal 2024,2 $3.206 billion a year later,2 and $5.389 billion in the first quarter of fiscal 2026.2 That is roughly three times the fiscal 2024 quarter, on our arithmetic. For an AI practitioner, this is what a memory shortage looks like on the supplier's side: the company that makes the memory is committing capital on a scale it did not a few years ago.
Micron's cash position moved too. The filings record $10.163 billion in the third quarter of fiscal 20253 and $24.995 billion in the third quarter of fiscal 2026.3 Cash was $13.908 billion in the second quarter of fiscal 2026.3 That is a sharp build in a single year, and it shows the spending is running alongside cash that is accumulating, not against a shrinking balance.
The cost side moved far less. Cost of revenue was $16.956 billion in fiscal 2023,4 $19.498 billion in fiscal 20244 and $22.505 billion in fiscal 2025.4 The first-quarter figure was $5.997 billion in fiscal 2026 against $5.361 billion in fiscal 2025.4 Capex roughly doubled in fiscal 2025 while cost of revenue rose by about a sixth, so the growth is going into capacity, not into the cost of what is currently sold. We offer that as our reading of the filings, not as a company statement. These filings do not tell us how much of the spending is aimed at AI memory specifically.
What analysts say, and how much weight it carries
Lynx Equity Strategies rates Micron "a clear buy with anticipated upside of 33%" and Sandisk "a clear buy with anticipated upside of 49%."5 These are one firm's forecasts, reported by a source where only 22% of 18 checked claims held up.5 They fit the bullish mood around memory names, but they are opinions, not data. Compare them with the SEC figures above, which are the part of this story that has been checked.
Dealmaking beyond memory: the network and the factory
Bull and Kalray announced a collaboration on networking technologies for AI and high-performance computing infrastructure, including compatibility with Ultra Ethernet, a new open standard.6 Kalray's Éric Baissus said the agreement "confirme une nouvelle fois la pertinence de notre vision technologique."6 In English, he said it confirms the relevance of Kalray's technology vision. The source's reliability rating is low: 31% of 2,928 checked claims held up.6
On the US side, GlobalFoundries signed a letter of intent with the Department of Commerce for a $300 million award to accelerate silicon photonics.7 A letter of intent is not a finished award, and the release says the money is "expected."7 AMD's Mark Papermaster said: "As AI systems scale, moving data efficiently is as critical as increasing compute performance."7 The point for practitioners is that moving data, not only computing on it, is now treated as a bottleneck.
Consolidation continues elsewhere. Skyworks and Qorvo announced the expected leadership of the combined company, effective on completion of a pending transaction.8 Bob Bruggeworth said the announcement reflects "the strong partnership that has shaped our integration planning efforts from the very beginning."8 This is not an AI-specific deal in the dossier, so we note it as part of the wider semiconductor consolidation picture and no more than that.
The custom-silicon push: ideas ahead of evidence
On October 8, Zeta introduced AthenaOS and Athena MCP, and announced the Athena Inference Model, developed with Fireworks using NVIDIA Nemotron.9 Zeta also said it has begun developing a specialised AIM chip with partners, through what it calls a capital-light approach.9 The announcement is notable because the chip is being pursued with partners and not built alone. The dossier gives no performance, cost or delivery figures for it, so we cannot judge it. The topic framing also names SambaNova and Cerebras as entrants moving toward custom inference chips, but the dossier contains no verified detail on either, and we do not add any.
A more radical idea comes from Hannah Earley, cofounder and CTO of Vaire Computing, who is building chips that recycle energy usually lost as heat, a strategy known as reversible computing.10 She compares conventional computing to "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."10 She also says: "I want to tackle every part of how computers are built, and rethink it in these terms."10 Be cautious with this source: none of the 11 claims checked from it held up (0%).10 The quotes are verbatim, but we cannot confirm the technical claims around them. It is an interesting direction, not an established one.
Small suppliers and insiders: weak signals
Solitron Devices reported net sales up 101% to about $5.44 million, with backlog up 28% to $23.34 million, while net bookings fell 48%.11 The company said it expects sales to continue "at this level or greater for the remainder of the 2027 fiscal year."11 It trades on the OTC Pink market and is very small, so it is a data point and not a trend. Separately, Lattice Semiconductor's SVP Esam Elashmawi reported selling 16,773 shares at a weighted average of $134.56.12 An insider sale is not on its own a signal about the company's prospects, and that source had 22% of 18 checked claims hold up.12
What we could not verify
The topic also references an FTC probe of AI labs, OpenAI's copying accusation against Moonshot, a DeepSeek-Huawei arrangement and espionage targeting AI experts. The dossier holds no source documents or verified facts on any of them. We therefore make no claims about them and treat them as unconfirmed risks.
What to watch
- Micron's next filings: whether capex keeps rising from the $5.389 billion first-quarter level, and whether cash keeps building from $24.995 billion.23
- Whether the Nvidia-SK Hynix co-development produces details beyond the stated aim of stable HBM supply.1
- Whether Zeta's AIM chip work produces specifications, partners or dates.9
- Whether the GlobalFoundries letter of intent becomes a finished award.7