AI Daily Brief: Nvidia Ships Its Groq Chip and a New Model of Its Own, Qualcomm Bets $14 Billion Against It

Today's news is really one story wearing four coats: the handful of companies that make AI chips are no longer content to just make chips. Nvidia shipped a model of its own this month. Qualcomm is spending billions to avoid ever needing Nvidia's blessing again. And Samsung's chairman is flying out to make sure Nvidia keeps needing him.

Nvidia's Groq chip heads into production, and Samsung's chairman heads to Jensen Huang's door

The first chip to come out of Nvidia's $20 billion Groq deal is about to ship. The Groq 3 LPU, a dedicated inference co-processor built around 512MB of on-chip SRAM, slots into the Vera Rubin platform as a decode-phase accelerator sitting alongside Nvidia's Rubin GPUs, with Nvidia claiming 35 times higher throughput per megawatt than Blackwell alone on trillion-parameter models. It ships in the third quarter, manufactured entirely by Samsung on a 4-nanometer process.

That manufacturing relationship is about to get closer. Samsung chairman Lee Jae-yong is expected to meet Jensen Huang in Silicon Valley by late July, their first meeting in nine months, to talk through HBM supply, advanced packaging, and Samsung's own AI data-center buildout at home. Samsung shares rose more than 5% on the news. Nvidia used to be a company that bought memory and foundry capacity from Samsung. Increasingly, it's a company whose entire inference roadmap runs through Samsung's fabs.

Qualcomm's $14 billion bet to not need Nvidia at all

Not every chip giant wants to get closer to Nvidia. Qualcomm is reportedly in talks to acquire Tenstorrent, the RISC-V AI chip startup run by veteran architect Jim Keller, for $8 to $10 billion, on top of an already-confirmed acquisition of Modular. Qualcomm shares jumped 4.3% on the report before giving back some of the move after hours, the kind of reaction that says investors like the ambition and aren't sure yet about the price tag.

Combined, the two deals commit more than $14 billion to a single goal: giving cloud providers and enterprise buyers a real option to run AI workloads on hardware that never touches an Nvidia part number. Tenstorrent's RISC-V accelerators target inference, the same workload Nvidia just spent $20 billion trying to own more of with Groq. Both companies are betting on the same growth curve. They disagree on whether Nvidia should get a cut of it.

The most interesting new model this week came from a chip company

While labs raced to ship the next flagship chatbot, Nvidia's own research team quietly open-weighted Nemotron-Labs-Diffusion, a 3B/8B/14B model family that can switch between ordinary autoregressive decoding and parallel diffusion decoding from the same set of weights. Run in its self-speculation mode, the 8B version decodes six times more tokens per forward pass than Qwen3-8B and roughly quadruples throughput on a GB200. The trick is architectural, not just scale: train one model on both objectives at once, and it learns to draft and verify its own output instead of needing a separate, smaller draft model bolted on the side.

It's a small release next to Kimi K3 or GPT-5.6, but it's a tell. The company that sells the hardware everyone else's models run on has decided the software layer is worth owning too.

What it means for India

Two of today's chip stories have an India angle, and both are about who does the actual engineering. Qualcomm's Bengaluru, Chennai and Hyderabad centers, its largest engineering footprint outside the US, already taped out a 2-nanometer design earlier this year. If the Tenstorrent deal closes, there's a reasonable chance some of that $14 billion bet against Nvidia gets designed, at least in part, out of India rather than San Diego.

Delhi, meanwhile, is running its own version of Qualcomm's playbook. The government's newly approved Semicon 2.0 program commits roughly ₹1.27 lakh crore, about $14 billion, to Indian chip-design startups, structured as government equity stakes alongside private venture capital rather than pure grants. The number lining up with Qualcomm's own spend is a coincidence, but the instinct isn't: everyone building AI hardware right now, a $180 billion company and a national government alike, has decided that owning the chip-design layer outright is worth a double-digit-billion-dollar bet.

Markets and AI money

	Detail
Qualcomm	shares up 4.3% to $220.81 on Tenstorrent talks, before giving back some gains after hours
Samsung	shares up more than 5% on news of the Lee Jae-yong / Jensen Huang meeting
Nvidia's Groq 3 LPU	targets $45 per million tokens, claims 35x throughput per megawatt over Blackwell alone
Nemotron-Labs-Diffusion	6x tokens per forward pass vs Qwen3-8B, roughly 4x throughput on a GB200

Every number in that table is really the same bet stated four different ways: that the next leg of AI economics gets won or lost on inference efficiency, not on whoever trains the single biggest model. Nvidia is betting it can own that shift by buying and building the co-processors itself. Qualcomm is betting it can own a slice of it by refusing to need Nvidia's permission. Samsung is betting it can be indispensable to both outcomes at once.

Three different bets on the same question: once training the biggest model stops being the moat, who actually owns the machine that runs it.