Closing Time: Google takes aim at Nvidia's moat

Nvidia’s most important advantage has never been a chip. It is the decade of software that makes AI code run best on its hardware, and on Tuesday Google took a swing at it. The company released TorchTPU, its long-awaited version of PyTorch, as an open-source project, TechCrunch reports.

PyTorch is the toolkit most AI developers use to build, train and run models, and by some accounts it sits under 85% of AI apps. Because it grew up on Nvidia’s GPUs, much of it was built to work best with them. That default is a big part of why the AI industry runs on Nvidia, even when other chips are cheaper or easier to get.

TorchTPU lets those same apps run on Google’s own AI chips, called Tensor Processing Units, and makes it easy to move existing PyTorch projects onto Google Cloud. Developers can make the switch with a simple configuration change, without rewriting their code.

Google is not alone. Amazon released its own version, TorchNeuron, on Monday, letting PyTorch apps move to its Trainium chips and AWS with no rewriting, and it is also submitting code to the main PyTorch project that loosens features from Nvidia’s hardware by default.

Google is trying to go a step further. It says it worked directly with Meta, Hugging Face, vLLM maker Inferact and SGLang maker RadixArk so TorchTPU behaves like the original PyTorch with the tools developers already use. It also launched a TPU agent suite to help developers move their apps over and fix errors along the way.

None of this means developers will abandon Nvidia overnight. But two of the biggest cloud companies are now making it nearly free to leave, which matters most to the startups and labs trying to stretch every dollar of compute.

Elsewhere in AI today:

Bloomberg reports the shine is coming off neoclouds, the companies that rent out data center computing for AI. South Korea’s Elice tanked 32% in its Seoul debut on Tuesday, weeks after Australia’s Firmus dropped its IPO plans, while Nscale faces pushback on its IPO valuation.

CNBC reports Salesforce agreed to buy customer engagement platform Braze in a deal valued at about $5.66 billion, deepening its bet on AI-driven marketing. Braze shareholders will receive $47.50 a share in cash, and the stock jumped about 50% on the news.

Bloomberg reports top commodity brokers are gearing up for compute futures, derivatives tied to the cost of renting the computing power used to build and run AI. The firms are preparing for possible approval from the Commodity Futures Trading Commission to trade them in the US.

Reuters reports China vowed to safeguard the rights of its companies after the US ordered civilian federal agencies to remove DeepSeek products and services built on its models. A foreign ministry spokesperson said China opposes the politicization of science, technology and trade issues.

---

Markets

The Dow did the heavy lifting on Tuesday, gaining about 1% on strong earnings from 3M and Northrop Grumman, Quartz reports. The S&P 500 rose about 0.3%, while the Nasdaq finished roughly flat. 43 of the 86 AI and tech stocks we track closed higher.

Software led our list as investors rotated out of chips, according to StockStory. MongoDB jumped 5.0%, a day after launching open-source agent templates, Adobe rose 3.9% and Tempus AI added 3.8%.

Chip equipment makers went the other way. Lam Research fell 3.2%, Arm slid 3.1% and KLA dropped 2.8%, with Eaton and Cloudflare also among the decliners.

Biggest winners

Biggest losers

Prices are Tuesday’s U.S. cash close, October 20, 2026 (vs. Monday’s close). Market data is a snapshot, not live.

---

Deals Desk

One compute deal worth watching.

Anthropic × Nscale — Oct 20, 2026

Lease of computing capacity at a second data center on Nscale’s Monarch campus · West Virginia, US

460 MW; the deal could be worth around $45 billion, following a first agreement at the same site in August

Status: In talks (reported) · Source: CNBC

Full tracker → https://www.theaigentic.com/deals