Tech rally drives stocks higher as AI trade is back on

Google jumped in after-hours trading while Nvidia and Advanced Micro Devices dropped on the heels of a report from The Information that has the search giant muscling in on the chip designers’ turf.

Per the report, Meta is in discussions with Google to spend “billions of dollars” to use its AI chips in the social media company’s data centers starting in 2027, and to begin renting access to Google chips from its cloud business next year.

Historically, Google has rented access to these chips through its cloud business rather than supply them directly to third parties. The report suggests that insiders believe a more direct foray could allow the company to grab a market share in chips amounting to about 10% of Nvidia’s annual revenue.

Google’s AI chips — TPUs, or tensor processing units — are having a moment. These semiconductors were used to train its latest genAI model, Gemini 3, which has received rave reviews, and are cheaper to use than Nvidia’s offerings. That’s sent the stock to record highs, surpassing Microsoft in market value along the way.

According to The Information, Meta is even mulling using TPUs for training, considered a much more demanding task, rather than just inference alone.

Shares of Nvidia and AMD, which sell GPUs for use in data centers, fell about 2% in postmarket trading, while Google gained around 2%.

During Nvidia’s conference call last week, CEO Jensen Huang was asked about the competitive threat posed by custom chips. He responded by talking up the difficulty of inference (“How could thinking be easy?”). That’s a not-too-subtle nod to the idea that his company’s GPUs will be the more effective solution compared to more cost-effective options. He also touted the company’s CUDA software as a selling point, because it’s more commonly used and therefore makes it easier for its buyers to go on and sell AI computing capacity.

Google has aimed to make its JAX software easier for developers over time by making its TPUs operable via open-source software tied to PyTorch (invented by Meta), overhauling how errors are reported, and introducing an extension that makes it easier to write custom code, among others.

When spectators saw Michael Jordan blossom as a basketball star, it made them want to buy Nike’s Air Jordans.

Mizuho reckons there’s a similar halo effect for Broadcom based on the rave reviews for Google’s latest GenAI model, which include this ringing endorsement from Salesforce CEO Marc Benioff.

“Gemini 3 was trained and powered on Google homegrown TPU chips, which benefits partner AVGO,” wrote Daniel O’Regan, Mizuho’s managing director of equity trading.

Custom chips (ASICs) are Broadcom’s specialty, and as O’Regan noted, Google and Broadcom codesigned these building blocks for Gemini 3. Boosts to Google’s capex budget have tended to buoy shares of Broadcom, since it’s a big beneficiary of these outlays.

The early positive reception to Gemini 3 implies that: a) Google will want to continue this relationship (and need more chips for training and inference!), and b) other GenAI developers might be more willing to pursue the custom chip route for AI models and inference, perhaps eating into market share for Nvidia’s GPU-based solutions.

To this end, Broadcom announced a collaboration with OpenAI in mid-October to develop and deploy 10 gigawatts of custom AI accelerators.

Nvidia CEO Jensen Huang has argued that GPU-centric data center solutions are superior because of how ubiquitous the firm’s CUDA software is in high-performance computing.