🔥 FuriosaAI Just Raised $620 Million — the Korean Chip Startup That Told Meta No

A year and a half ago, a Seoul chip startup looked at an $800 million buyout offer from Meta and walked away (TechCrunch). That looked either brave or reckless, depending on which venture capitalist you asked.

Here’s the receipt. FuriosaAI just closed a $620 million Series D at a valuation of about $2.9 billion (The Next Web). That’s roughly four times the $735 million it was worth a year ago, and more than three times what Meta tried to pay for the whole company.

The most eye-catching line in the cap table isn’t a Silicon Valley fund. It’s a government. South Korea’s National Growth Fund put in $280 million — the single largest check in the round (The Next Web).

When a national government writes the biggest check to a chip startup, the money is making a point as much as a bet.

🧠 Why This Matters

Almost every dollar of the AI boom has run through one company’s silicon. Nvidia sells the chips that train the models, and it sells the chips that run them afterward. FuriosaAI is going after the second half — inference, the stage where a trained model actually answers your question — and it’s doing it from outside the United States.

That’s the part that turns a funding round into a geopolitics story. The chips powering Korea’s AI industry are, overwhelmingly, imported American ones. A $280 million state-backed check into a domestic alternative is Seoul trying to keep some of that supply at home.

For you, the reader, the signal is simpler: the market now believes there’s real money in not buying from Nvidia — enough that saying no to Meta looks less like ego and more like timing.

📊 Deep Dive

FuriosaAI was founded in 2017 by June Paik, a former Samsung and AMD engineer (TechCrunch). Its current chip, RNGD (pronounced “renegade”), went into mass production this past January (The Next Web).

The pitch is efficiency. Here’s how RNGD stacks up against the usual approach:

• Power draw: RNGD pulls 180 watts and runs on air cooling — no liquid-cooling plumbing required. Many top-end AI accelerators need far more power and elaborate cooling.
• Throughput: FuriosaAI claims RNGD matches the token throughput of Nvidia’s RTX PRO 6000 card-for-card (The Next Web).
• Efficiency: The company says it serves roughly twice as many users per kilowatt of power — the metric that actually moves data-center economics.
• Volume: TSMC is building about 20,000 RNGD units in 2026, with capacity secured for more than 40,000 in 2027 (The Next Web).

The customer list is early but real. Samsung SDS launched a cloud service running on RNGD in July. LG runs its EXAONE models on the chips. Korean AI firm Upstage uses them for search summaries, and Equinix brought RNGD to Europe through Lisbon (The Next Web).

The round pulled in an unusually global crowd behind the government money: $110 million from international investors including Infinitum, Jericho Capital, Cisco Investments and Abu Dhabi’s NeoFort Holding, plus $125 million from existing shareholders NAVER, DSC Investment and KDB Capital (The Next Web).

“Taking an advanced AI chip from design to volume production is an exceptionally difficult undertaking.” — John Yetimoglu, Chief Investment Officer, Infinitum (The Next Web)

⚠️ The Catch

Those performance numbers are FuriosaAI’s own, measured by FuriosaAI, and no independent lab has confirmed them. Matching a specific Nvidia card on throughput in a chosen benchmark is not the same as matching Nvidia across the messy range of real production workloads.

And raw silicon was never Nvidia’s only moat. CUDA — the software layer developers have built on for nearly two decades — is the stickier lock-in. A chip that’s cheaper per watt still has to fit into toolchains that assume Nvidia at every layer. That’s a migration cost measured in engineering months, not spec sheets.

Then there’s the uncomfortable math of 20,000 units a year. Nvidia ships data-center GPUs by the millions. FuriosaAI isn’t trying to replace Nvidia; at this volume it’s fighting for the slice of workloads where efficiency matters more than raw peak performance.

“We see strong demand and opportunities as we are scaling our business globally.” — June Paik, CEO, FuriosaAI, on Bloomberg Television (The Next Web)

🎯 What Happens Next

The next chip is already in the lab. FuriosaAI’s third-generation accelerator is being built with Broadcom on a 2-nanometer process, with a multi-die design and HBM4 memory (The Next Web). Broadcom’s involvement matters — it’s the company quietly co-designing custom AI silicon for several hyperscalers.

Watch two numbers through 2027. First, whether that 40,000-unit capacity actually fills with paying orders rather than pilots. Second, whether any customer outside Korea commits at scale — Paik says talks are underway across Asia-Pacific and the Middle East, and that’s exactly where a non-Nvidia option gets attractive.

🧩 Bigger Picture

FuriosaAI is one data point in a much larger pattern: governments have decided AI chips are infrastructure, not just products. Korea’s National Growth Fund here, sovereign money from the Gulf in the syndicate, export-finance banks writing their first startup checks — the capital stack around advanced silicon is starting to look like the one around power grids and ports.

The inference market is the prize everyone’s circling because it’s the part that never stops. You train a model once; you run it billions of times. Whoever makes running it cheaper per watt gets paid on every single query, forever.

Meta offered $800 million for a shortcut to that market and got turned down. The company that said no is now worth $2.9 billion and building its third chip. Sometimes the best deal is the one you don’t sign.

---

Sources

• The Next Web — FuriosaAI raises $620m at a $2.9bn valuation for its Nvidia rival chips
• TechCrunch — AI chip startup FuriosaAI reportedly turns down $800M acquisition offer from Meta
• Tech Startups — Startup Funding News Today, October 29, 2026

❓ Quick answers

How much did FuriosaAI raise and at what valuation?

FuriosaAI raised a $620 million Series D at a valuation of about $2.9 billion, announced October 28, 2026 (The Next Web).

Did FuriosaAI really reject Meta?

Yes. In March 2025 FuriosaAI turned down an approximately $800 million acquisition offer from Meta, reportedly over post-acquisition strategy rather than price (TechCrunch).

What is the RNGD chip?

RNGD is FuriosaAI's second-generation AI inference accelerator, a 180-watt air-cooled card that entered mass production in January 2026, with about 20,000 units being built by TSMC in 2026 (The Next Web).