Between October 10 and 12, the Wall Street Journal first reported — and multiple Chinese outlets confirmed — that Amazon is in talks to acquire the AI startup Decart for roughly $7 billion. If signed, it would become another landmark deal in the cloud giants' compute race. As of publication, no final agreement has been reached and terms may still change.

From Foundation Models to "Chip-Switching": What Makes Decart Scarce

Unlike celebrity startups chasing frontier models, Decart's core asset is consistently described across outlets as AI chip-switching software — tooling that lets AI training and inference workloads migrate and run across accelerator chips from different vendors. Reports by e-Nanyang and Phoenix Tech both stress that this capability is exactly what Amazon wants: when GPU supply is tight and home-grown silicon is ramping up, whoever controls the switching layer holds the pricing power.

When compute becomes an arms race, "escape from lock-in" itself becomes an asset you can acquire.

The Window After Anthropic Walked Away

Phoenix Tech reported that Anthropic, which had previously engaged with Decart, ultimately walked away, and Amazon promptly became the most active suitor. For Amazon — a major Anthropic backer that also develops its own Trainium chips — the deal would shore up its software layer while reducing dependence on a single GPU ecosystem. Sohu's commentary framed it as "Amazon's AI compute anxiety and the global infrastructure arms race."

Deal Highlights (as reported)

· Size: about $7 billion; talks ongoing, not yet signed

· Target: Decart, known for AI chip-switching software

· Buyer: Amazon; Anthropic had engaged earlier and passed

· First reported: around October 10 (Wall Street Journal)

The New M&A Logic: Buy Compute, but Above All Buy Freedom from Lock-in

Place the deal in its wider context: Tencent plans an offshore bond to fund capex, NVIDIA raised its buyback authorization to $235 billion, and Broadcom expects a major jump in AI chip sales. In this race, what giants fear most is not a shortage of cards but being bound to a single ecosystem. Chip-switching software turns a ten-thousand-GPU cluster from a sunk cost into a negotiable asset.

The lesson for enterprises is equally direct: heterogeneous compute scheduling is rising from an engineering detail to a strategic capability. NineZenith's Zenith-Act industry LLM platform supports unified scheduling across models and chips in government and enterprise scenarios, keeping client workloads continuous through supply fluctuations — the same underlying logic as Amazon's reported acquisition.

(Compiled from public reports by The Wall Street Journal Chinese, EO (Economic Observer), Phoenix Technology, Eastmoney, e-Nanyang and others)