In one day in San Francisco in late July, roughly $950 billion in new artificial-intelligence agreements were signed.1 Jensen Huang was in the room. So were the heads of Samsung, SK Group, Hyundai Motor and Naver, there for a summit convened around South Korea's president. By the end of it, Nvidia had used the moment to lock down something more specific than a headline number: a deal to help secure its supply of the memory chips its AI processors depend on.1
Why memory, and why now
Nvidia's most advanced AI chips are only as fast as the memory sitting next to them — a specialized product called high-bandwidth memory, or HBM, made almost exclusively by a handful of South Korean and Japanese firms. Nvidia's expansion with SK Hynix goes further than a purchase order: according to Raj Mirpuri, discussing the arrangement, "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."1 That is Nvidia embedding itself in its supplier's product roadmap, not just buying what's on the shelf — a sign of how tight the memory market has become.
That tightness is showing up in how investors are pricing memory-chip makers. On September 4, ASML — the Dutch company whose machines print the circuits onto virtually every advanced chip made anywhere — rose more than 4% in a single session on no news of its own.2 The move was a sympathy rally, triggered by a bullish research note on the memory sector from a firm called Lynx Equity Strategies, which rated Micron Technology a "clear buy" with 33% anticipated upside and Sandisk a "clear buy" with 49% anticipated upside.2 If you hold a broad U.S. stock index fund, there is a good chance you already own a slice of these companies — ASML, Micron and Sandisk are all widely held large-cap names, though the dossier here does not specify exact index weightings, so treat that as a general observation rather than a precise figure. Read the actual size of that bet cautiously: it is one research firm's call, not a verified outcome, and the dossier does not show how that firm's past calls have performed.
The infrastructure race is also a government bet
The memory story is playing out alongside a parallel push to build faster connections between chips. On July 29, GlobalFoundries signed a letter of intent with the U.S. Department of Commerce under which its CHIPS Research and Development Office is expected to award the company $300 million to develop next-generation silicon photonics — technology that moves data between chips using light instead of electrical signals.3 AMD, a customer for that kind of packaging technology and a direct competitor to Nvidia in AI chips, welcomed the move. Mark Papermaster, AMD's representative, put the stakes plainly: "As AI systems scale, moving data efficiently is as critical as increasing compute performance. 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."3
The same week, a similar dynamic played out in Europe: French firms Bull and Kalray announced a partnership to develop next-generation high-speed networking for AI and high-performance computing infrastructure, including compatibility with Ultra Ethernet, an emerging open networking standard for future supercomputers.4 Kalray's chief executive, Éric Baissus, framed the deal as confirmation of the company's technology strategy and said the collaboration aims to build the networking layer the next generation of AI systems will need.4 Individually, neither of these is a headline-grabbing figure. Together, they show governments and mid-sized chip suppliers on both sides of the Atlantic treating the physical plumbing between chips — not just the chips themselves — as strategically important enough to fund directly.
The boom reaches small suppliers too
The clearest sign that this demand is broad rather than confined to giant companies comes from one of the smallest names in the dossier. Solitron Devices, a Florida-based maker of semiconductor components, reported net sales roughly doubling — up 101% to about $5.44 million in its fiscal first quarter, versus $2.70 million a year earlier — with backlog up 28% to $23.34 million.5 The company said it expects revenue to continue at this higher level or better for the rest of its fiscal year, "driven by strong backlog."5 A company this size does not move markets, but it is a useful gauge: when demand is strong enough to double order books at a niche supplier, it suggests the AI chip cycle is not confined to the handful of giant names that dominate the headlines. That said, this comes from a single company announcement carried by a source where, across checks of thousands of its claims, only about three in ten held up — a reminder to treat the specific figures as reported rather than independently verified.
A more speculative frontier: chips that waste less energy
One more thread worth flagging, with a clear caveat. A UK startup called Vaire Computing is working on "reversible computing" — chips designed to recycle the energy conventional processors throw off as heat, rather than dissipating it. Its co-founder, Hannah Earley, describes the ambition in blunt terms: conventional computing, she says, "is like racing through a city only to pump the brakes at every intersection: the car loses momentum and must burn more fuel to accelerate again."6 She wants to rebuild chips from the ground up around that principle: "I want to tackle every part of how computers are built, and rethink it in these terms."6 This is an early-stage, unproven approach, not a near-term product — and it's worth being explicit that the single source describing it, in checks of its recent claims, had none hold up verification. That doesn't mean the reporting is wrong, but it means readers should treat this as a story to watch rather than a verified technology.
What to watch
The clearest, best-sourced fact in this cycle is the scale of dealmaking around Nvidia's supply chain — the $950 billion in agreements struck in a single day and the SK Hynix memory arrangement built to survive beyond a single purchase order.1 The memory-chip upside calls from Lynx Equity Strategies are real analyst opinions that moved ASML's stock, but they are one firm's forecast, not a settled outcome — watch whether Micron's and Sandisk's actual results over the coming quarters bear them out.2 And the government-backed investments in silicon photonics and AI networking, in the U.S. and in France, are early-stage commitments whose payoff — faster, more efficient connections between chips — is still being built, not yet delivered.34
