Meta eyeing AI chip startup FuriosaAI

Nvidia GPUs are expensive, scarce, and driving Big Tech's power bills through the roof. Meta's latest response? Just buy the company making cheaper, cooler alternatives.

Meta is in advanced talks to acquire FuriosaAI, a South Korean fabless chip startup, according to a report from Forbes. The deal could be finalized as early as this month, though Meta is reportedly not the only tech giant kicking the tires.

Before the acquisition talks surfaced, FuriosaAI was in the middle of raising fresh funding at an 800 billion won ($550 million) valuation. Founded in 2017 by former Samsung engineer June Paik, who also worked at AMD, the startup has raised $115 million to date from backers like Naver and DSC Investment.

Half the performance, a quarter of the power

While most chip startups try to challenge Nvidia on brute training speed, FuriosaAI took a radically pragmatic angle: inference efficiency.

In August 2024, the startup launched RNGD, an accelerator built specifically for hyperscale data centers and large language models like Meta's Llama family. Paik noted last fall that single H100 GPUs were commanding around 50 million won while facing severe supply shortages.

According to Paik, RNGD delivers about half the raw performance of Nvidia's H100, but sips just a quarter of the electricity. Total system build-out costs drop by roughly 50% compared to equivalent Nvidia setups.

For a company like Meta, that math makes total sense. Meta expects its capital expenditures to hit up to $65 billion this year, with the vast majority earmarked for AI infrastructure. Power constraints, not just silicon supply, are quickly becoming the hard ceiling for scaling data centers.

Meta has already been developing in-house accelerators alongside custom silicon with Broadcom. Snapping up a team that already has a working, power-sipping LLM chip simply speeds up the timeline.

Why it matters

If you build with open-source AI, this deal points directly toward cheaper tokens. Meta footing a massive portion of the world's Llama inference bill means any hardware cost reduction directly trickles down into ecosystem sustainability.

For the semiconductor industry, this signals that the M&A window for inference chip startups is wide open. Building cutting-edge GPUs from scratch takes years, and hyperscalers do not have the luxury to wait. Buying specialized fabless teams is the quickest path to chip independence.

We think Meta's approach is remarkably shrewd. Taking on Nvidia in front-line training is a money pit, but slashing power and dollar costs in inference is where real balance-sheet relief happens. RNGD's trade-off—giving up top-end speed in exchange for drastic energy savings—is tailor-made for Meta's massive deployment footprint.

Still, integration is never frictionless. Meta will have to assimilate FuriosaAI's proprietary software stack into its existing PyTorch-heavy infrastructure without losing engineering momentum. With other suitors also circling, the deal is not done yet, but the intent is crystal clear.