A $250M Chip Bet With No Physical Chip to Show For It

The Deal That Shouldn’t Exist

In May, The Information reported that Anthropic was in talks to buy $250 million worth of chips from a British startup called Pixel. Here is the problem: the chip does not physically exist. It is still in the design phase. A first article is not expected until next year. No proof-of-concept data. No samples. No wafer runs.

This is not a normal deal. In a sector where data-center operators spend fortunes measuring silicon performance before signing contracts, Anthropic is buying silicon that has never left a drawing board. The amount—roughly 350 billion won—would make it one of the largest pre-production chip deals in recent memory.

The startup behind it, Pixel, was founded in 2022 by Walter Goodwin, an Oxford robotics PhD. It has about 100 employees. After the Anthropic report surfaced, its valuation jumped to $6.5 billion. It is now reportedly preparing to raise another $600 million. A hundred-person company, a six-figure burn rate, and half a billion dollars in fresh capital, all chasing a chip that only exists on paper.

The Memory Wall Is Getting Worse

To understand why anyone would make this bet, you have to understand the single biggest bottleneck in modern AI: the distance between memory and computation.

Every large language model is, fundamentally, a program that stores massive amounts of data in memory and moves it to processing units for computation. The more parameters a model has, the more data it must shuttle between DRAM stacks and GPU compute cores. That movement is the bottleneck. It is slow. It consumes enormous bandwidth. It generates heat. It costs money.

The industry’s answer has been High Bandwidth Memory—HBM. Samsung Electronics and SK Hynix produce it. It stacks DRAM vertically and connects it directly to GPUs through advanced packaging. It works. It is also getting more expensive, more power-hungry, and more thermally challenging every generation.

Pixel claims its “in-memory” approach—what the industry calls PIM, or Processing In Memory—could solve this by embedding computation directly inside the memory array. Move the logic into the memory, and you eliminate the data movement that creates the bottleneck. Pixel says its approach achieves 25 times faster inference at one-tenth the cost of comparable HBM-based solutions. The numbers sound almost too good to be true. They probably are. But the direction is right.

Samsung’s Surprise Reveal

Samsung has been working on PIM for years. The company described its approach publicly for the first time at Hotchips, the year’s most important applied semiconductor conference, held in late August in Colorado. The reveal was notable because Samsung typically keeps PIM research behind closed doors until it has a shipping product. Announcing it this early—when other players like Google and Microsoft are quietly developing their own in-memory approaches—suggests Samsung feels pressure to signal relevance in a race it might otherwise miss.

The competitive landscape is shifting. TSMC controls the leading-edge fabrication nodes that make advanced AI chips possible. Samsung foundry is a distant second, and its memory business—once the crown jewel—is under strain from HBM pricing wars with SK Hynix. If PIM becomes the next architecture shift, as several industry observers believe it could, Samsung’s memory expertise gives it a structural advantage over pure-play compute companies. But expertise is not execution. Pixel’s claim that a 100-person team solved what Samsung’s thousands of engineers have not is exactly the kind of disruption story that makes investors nervous and competitive teams wake up at 3 AM.

Who Wins, Who Loses

If Pixel delivers—even at a fraction of its promises—it upends the HBM value chain. SK Hynix and Samsung would face demand compression on their most profitable product line. TSMC’s dominance in advanced packaging gains relevance if in-memory architectures require new interconnect standards. Anthropic’s bet is a hedge: if PIM works, it locks in supply before anyone else can. If it fails, it loses $250 million and looks foolish. The asymmetric upside is the whole point.

For the United States, the implications are indirect but real. Anthropic’s purchase goes to a British startup, not a Chinese company, so US export controls do not block it. But if PIM becomes the dominant architecture for AI training and inference, the geopolitical dynamics of chip supply chains shift again—this time away from fabrication node dominance toward memory integration expertise.

South Korea’s position is the most complicated. Samsung and SK Hynix have built their fortunes on memory. PIM threatens to make their current products less central to AI systems. At the same time, Samsung’s early PIM announcement shows the company is not standing still. The question is whether it can translate research labs into shipping silicon fast enough to stay ahead of startups like Pixel, which move with no institutional overhead and every incentive to surprise.

The Real Story

The real story here is not a single deal. It is the direction of the industry. Every major AI company is treating the memory wall as an existential threat. HBM is the current answer, but it is getting harder to scale. In-memory computing is the emerging alternative. The race is early. The bets are large. And the first company to ship a working PIM chip at scale will have an architectural advantage that will be extremely difficult to displace.

Anthropic is not buying a chip. It is buying optionality on the next architecture. That is a much more interesting proposition—and a much riskier one.