The A.I. Race Spreads Beyond the Chip

The A.I. Race Expands Beyond Chips

The artificial intelligence boom is no longer defined only by the scramble for Nvidia’s graphics processors. This week, a new set of developments made clear that the contest is spreading across nearly every layer of the technology stack — from how data moves inside machines, to who controls memory supply, to whether start-ups can finance and build infrastructure on a scale once reserved for the largest cloud companies.

Nvidia has moved deeper into photonics, placing multibillion-dollar bets on optical technology that uses light, rather than electricity, to shuttle data through A.I. systems. Mistral, the French start-up, said it was exploring designing its own chips as it builds out data-center capacity in Europe. Samsung began shipping samples of its next-generation high-bandwidth memory, a crucial component for A.I. servers. And Anthropic, one of the leading model makers, raised another enormous financing round that valued it at nearly $1 trillion.

Taken together, the moves suggest that the A.I. buildout is broadening and becoming even more capital-intensive. What began as a battle over access to top-end processors is turning into a wider struggle over interconnects, memory, power, cloud capacity and the money needed to secure all of them.

A New Bottleneck: Moving Data Fast Enough

For years, much of the industry’s focus was on compute — how many powerful chips a company could acquire and link together. But as A.I. systems have grown larger, another problem has become harder to ignore: moving data between those chips quickly and efficiently enough to keep them working at full capacity.

That is where photonics has drawn fresh attention. In March, Nvidia said it would invest $2 billion each in Lumentum and Coherent, suppliers working on optical technologies that can help move data using light. The deals included not just investments but purchase commitments and future access rights, a sign that major buyers are trying to secure supply before the technology becomes a more common part of A.I. infrastructure.

The attraction is straightforward. Optical interconnects are seen as a way to cut power consumption and reduce latency at a time when giant A.I. clusters are straining the limits of electricity use and data transfer. If chips are the engines of A.I., photonics could become part of the plumbing that allows those engines to run at scale.

Whether the technology will move quickly from strategic bet to standard deployment remains uncertain. But Nvidia’s willingness to put billions behind it shows how seriously the company views the problem.

Start-Ups Seek More Control

The push is not coming only from established suppliers. Model developers, too, are trying to gain greater control over the infrastructure beneath their products.

Mistral, which has positioned itself as Europe’s answer to American A.I. leaders like OpenAI and Anthropic, said it wanted a more “full-stack” role and was exploring chip design as it expands its infrastructure footprint. The company has said it is targeting 1 gigawatt of electrical capacity by 2030 and that its infrastructure plans represent roughly €4 billion in investment.

That ambition reflects a shift now underway across the sector. A.I. start-ups that once relied primarily on outside cloud providers are increasingly looking to own or shape more of the underlying hardware and data-center architecture. The logic is partly strategic and partly economic: dependence on a handful of suppliers can leave companies exposed to shortages, pricing power and geopolitical risk.

For Mistral, the effort also carries a regional dimension. Europe has long worried about lagging the United States and China in advanced computing infrastructure. Building more of its own stack could give the company a stronger claim to technological sovereignty. But it also raises difficult questions about execution. Designing chips is expensive and slow, and there is no guarantee that in-house hardware will produce a clear competitive edge in model performance or cost.

Memory Remains a Critical Choke Point

Even if processors remain the most visible symbol of the A.I. era, memory has become one of its most stubborn bottlenecks.

Samsung said it had begun shipping samples of its 12-layer HBM4E memory chips to customers around the world, sending its shares sharply higher. The company said the new product was more than 20 percent faster than its prior HBM4 generation.

High-bandwidth memory, or HBM, has become indispensable in A.I. servers because it allows processors to access data at very high speeds. Without enough advanced memory, even the best A.I. chips cannot perform efficiently. That has turned HBM into one of the industry’s most strategically important components and intensified competition among Samsung, SK Hynix and Micron.

Samsung is trying to regain ground after rivals, particularly SK Hynix, established strong positions in supplying memory for A.I. systems. Shipping samples is only one step; the more consequential test will be whether Samsung can ramp production at scale and win meaningful orders from major customers. Still, the market reaction reflected how eager investors are for any sign that supply constraints in A.I. hardware may ease — or that another winner may emerge.

Financing on an Infrastructure Scale

If the hardware race is expanding, so is the financing needed to stay in it.

Anthropic’s latest $65 billion round, which valued the company at about $965 billion post-money, underscored the degree to which leading A.I. laboratories are now being treated less like software start-ups and more like infrastructure platforms that consume extraordinary amounts of capital. At that valuation, Anthropic would eclipse OpenAI as Silicon Valley’s most highly valued A.I. start-up.

The size of the round points to a basic reality of the current market: capital intensity is still rising, not falling. Training and serving cutting-edge models requires chips, data centers, power contracts and engineering talent in volumes that few companies can finance from operations alone.

For investors, the wager is that a small number of frontier model companies will eventually occupy commanding positions in the next computing era. But the numbers also sharpen the doubts. It remains unclear whether valuations approaching the trillion-dollar mark can ultimately be justified by durable profits, or whether they are being propelled mainly by scarcity, momentum and the fear of missing out.

Investors Look Beyond Model Makers

Public markets are also beginning to reward companies connected to A.I. deployment, not just those building the models themselves.

Okta rose after reporting stronger quarterly results and pointing to demand around “agentic” A.I. — systems capable of taking actions on a user’s behalf. The company said it was allocating more resources to tools aimed at securing such systems, a sign that the spread of A.I. is creating demand for a supporting layer of identity and security software.

Nebius, a Dutch cloud provider, also surged after a fund run by Leopold Aschenbrenner, a former OpenAI researcher, disclosed a sizable stake. The move highlighted a growing enthusiasm for cloud and infrastructure names seen as adjacent beneficiaries of the A.I. spending boom.

Those gains suggest that the market is widening its definition of an A.I. company. It is no longer only the chip designers and model labs drawing attention, but also the businesses that can help deploy, host, secure or accelerate A.I. workloads.

Why This Moment Matters

The new wave of announcements marks a turning point in the shape of the A.I. race. The first phase was dominated by GPUs. The next is about the rest of the machine: memory that can feed those processors, optical links that can connect them, power and data centers that can sustain them, and financing large enough to pull the whole system together.

That shift matters because it changes who can compete. Success will depend not only on designing a better model, but also on securing scarce industrial inputs and raising vast sums of money. It also means that the winners may come from unexpected corners of the market — component suppliers, cloud operators, security firms and companies with control over physical infrastructure.

Yet many of the central questions remain unresolved. It is not clear how quickly photonics will become mainstream in A.I. systems, whether Samsung can close the gap with established HBM leaders, or whether Mistral can turn infrastructure independence into a commercial advantage. And while investor enthusiasm remains strong, it could fade if revenue from A.I. applications fails to catch up with the breathtaking pace of spending.

For now, though, one conclusion is becoming harder to dispute: the A.I. boom is not narrowing. It is spreading outward, into every expensive and strategic layer required to make the technology work.