A strategic acquirer and a venture syndicate are not bidding on the same object, which is worth holding on to before reading the distance between the two MatX numbers as a markdown. The roughly $7bn Reuters describes would have bought control and the two people who between them ran the AI software for Google's TPUs and led the design of the hardware [1][6]; the roughly $4bn MatX is now reportedly seeking buys a minority position in a company whose first chips are slated for TSMC production and 2027 shipment [3][8]. Different goods. The gap is still $3bn, or 43% off the discussed price, and Reuters says it could not establish what ended the talks [1][4].
Then the scale test, because affordability is the explanation people reach for first. Anthropic committed more than $100bn to AWS over ten years in April for up to 5 gigawatts, which averages better than $10bn a year, so the whole MatX price was about eight months of one cloud contract [11][2]. Against the $65bn annualized run rate reported at the end of July, $7bn is under 11% [10][3]. Against the roughly $250m of inference silicon Anthropic has preliminarily agreed to buy from Britain's Fractile, it is 28 times the size [14][4]. A company that went from nearly $9bn of run rate at the close of 2025 to $47bn in May to $65bn in July, 7.2x in about seven months, had the cash on hand to buy a design house outright [10][5]. Whatever killed the talks, it was not the cheque.
The gap reads three ways, depending on what you think diligence found. Perhaps the pass was informational: the roadmap did not survive inspection, and $4bn is the market moving toward the asset. Or $4bn is simply an ordinary venture price, and $7bn was a control premium with retention packages stapled to it, in which case only the headline has been repriced. Or Anthropic decided that a collaboration, which is where one Reuters source says the discussions now lean, delivers the engineering access without consolidating a pre-revenue schedule onto its own roadmap [5].
This is probably wrong, but the third reading fits the surrounding behaviour better than the first. Anthropic said on 5 August that it was building its own chip engineering team while continuing to buy from Amazon, Google, Nvidia and AMD [13], and it already trains and serves Claude across TPUs, Trainium and GPUs at once [15]. MatX aims at training and reinforcement learning throughput [9], which is where Anthropic's largest cost sits, and a 2027 delivery date means two more years of buying someone else's parts whoever owns the design team [8]. Meanwhile Amazon has $5bn invested and as much as $20bn more pledged, up to $25bn, on the other side of that same relationship [12][6].
What would break the thesis: if MatX closes at or above $4bn with a strategic name on the round, the down-round reading collapses and this was simply two buyers pricing two different instruments. If it closes lower, or slowly, diligence was the story all along.
