FuriosaAI Secures $620M Series D to Scale RNGD Inference Chips

Direct answer: FuriosaAI, the South Korean AI chipmaker, has raised $620 million in a Series D funding round, valuing the company at approximately $2.9 billion. The round, the largest ever secured by a South Korean startup according to Korean reporting, will fund a major expansion of production for its RNGD inference accelerator and development of a third-generation chip. For UK small businesses, the significance is indirect but real: the cost of running AI, not just building it, is becoming the decisive battleground, and every credible challenger to the current hardware order puts downward pressure on the prices small firms eventually pay for AI features.

What happened

FuriosaAI announced the completion of its Series D on 29 October 2026 in South Korea, which is 28 October in US and UK time. The round totals $620 million, or 820 billion won, and brings the company’s post-money valuation to about $2.9 billion, roughly four times its previous level according to Bloomberg’s reporting of the deal. The company had begun fundraising in January 2026 with a target of between 450 billion and 750 billion won, and strong demand from domestic and international investors pushed the final total above that range.

The investor list is unusually broad, and it tells a story on its own. South Korea’s National Growth Fund was the largest investor, committing $280 million, in line with the country’s push to build domestic AI computing infrastructure and strengthen its semiconductor ecosystem. International investors contributed $110 million in total. The largest international investor was Infinitum, a US investment firm with deep experience in AI semiconductors that previously co-led a $650 million round for Groq, another inference-focused chip company. Singapore-based QUDA Capital Partners also joined, along with Cisco Investments, Abu Dhabi-based NeoFort Holding, and US firms Jericho Capital and Worldline Capital.

Existing shareholders contributed $125 million in follow-on investment. All three of FuriosaAI’s original seed investors, NAVER, DSC Investment and KDB Capital, took part again; DSC Investment has backed the company in every round since its seed investment in 2017. The Korea Development Bank added a further $22 million, while Nextrans and Kakao Investment, which joined in a Series C bridge round last year, increased their commitments. New Korean institutional investors also joined, including the Export-Import Bank of Korea, which selected FuriosaAI for its first direct investment in a startup, and United Asset Management, which committed $30 million, the largest commitment among new domestic investors after the National Growth Fund.

Some Korean outlets convert the round slightly differently, putting it at about $570 million and the valuation at about 3.8 trillion won, or roughly $2.6 billion, depending on the exchange rate used. The difference is a currency conversion effect, not a disagreement about the won figures, which are consistent at 820 billion won raised and about 3.8 trillion won post-money. This article uses the company’s own dollar figures, $620 million and about $2.9 billion, and notes the won totals alongside them so readers can reconcile the two.

FuriosaAI was founded in 2017 in Seoul by June Paik, and it now has offices in the United States, Portugal and Singapore as well as its Seoul headquarters. That footprint matters: this is not a company that intends to serve only the Korean market. Its chief executive describes the Series D as the moment the company moves from proving its technology to proving its business at global scale, with its chip already in volume production and in deployment with customers.

How the technology actually works

To understand why this funding round matters, it helps to separate two different jobs in AI computing: training and inference. Training is building a model, feeding it vast amounts of data and adjusting its parameters until it performs well. It is enormously expensive, it happens relatively rarely for any given model, and it is dominated by the most powerful hardware available. Inference is using a trained model: answering a question, summarising a document, transcribing speech, generating an image, or powering a chatbot’s reply. Inference happens billions of times a day, and its cost accumulates with every single use.

For a small business, inference is the cost that matters. You do not train your own foundation model. You pay, directly or through a software subscription, for someone else’s model to run on your behalf, thousands or millions of times. If inference gets 20% cheaper, the AI features in your customer service tool, your bookkeeping software, your marketing platform and your search tool all get cheaper, or better for the same price, or both. If inference stays expensive, AI features stay rationed: limited usage caps, higher subscription tiers, and features held back from smaller customers.

FuriosaAI’s bet is that inference deserves its own hardware, designed from scratch for that job, rather than borrowing hardware designed for training or for graphics. Its chips are built on what it calls the Tensor Contraction Processor, or TCP, architecture. Most AI accelerators, including GPUs, break model operations into fixed-size, two-dimensional matrix multiplications. FuriosaAI instead makes a higher-dimensional operation, tensor contraction, the basic unit of work for both hardware and software. Combined with reconfigurable compute slices and deterministic rather than dynamic execution on the silicon, the company says this reduces energy-wasting data movement while keeping the flexibility to support new models and new optimisation techniques.

The practical claim is concrete and testable. FuriosaAI says its 180-watt, air-cooled RNGD card delivers throughput comparable to a GPU drawing twice as much power or more. Power is not a footnote in data centres; it is often the binding constraint. A data centre has a fixed power budget and a fixed cooling capacity, and a chip that delivers more useful work per watt lets an operator serve more customers from the same building. That is why FuriosaAI emphasises sustainability alongside performance: more tokens from a fixed power budget is both a cost argument and an environmental one.

On the software side, FuriosaAI has tried to lower the barrier for developers. Models can be loaded from Hugging Face, defined in PyTorch, and run through a drop-in replacement for vLLM, a widely used system for serving large language models. Developers who need finer control can drop down to FuriosaAI’s Tensor Contraction Language and a low-level virtual instruction set. That combination, familiar tools on top, specialist control underneath, is important, because a chip that requires every customer to rewrite their software stack will struggle no matter how good its benchmarks are.

RNGD, the second-generation accelerator, entered mass production in January 2026. It uses high-bandwidth memory and is designed to fit within existing data centre power and cooling constraints, which means operators do not need to rebuild their facilities to adopt it. FuriosaAI says it expects to produce 20,000 RNGD accelerators in 2026, with plans to more than double production to over 40,000 units in 2027. Those are still small numbers compared with the millions of GPUs shipped each year, but the growth rate, and the fact that production is already at volume rather than at prototype stage, is what the Series D is funding.

The deployments named by the company give a sense of where RNGD is already working. LG Group is running its EXAONE deployments on RNGD. Upstage, a Korean AI company, now uses RNGD to serve live search summarisation for millions of users. Samsung SDS launched a commercial service, described as NPU-as-a-service, on RNGD in July 2026. None of these is a household name in the UK, but together they show a pattern: real, paying, production workloads, not just benchmark scores.

Comparison: how FuriosaAI sits against Nvidia and the other challengers

No honest discussion of AI chips can avoid Nvidia, which dominates AI computing to a degree rarely seen in technology markets. Nvidia’s GPUs train most frontier models, its CUDA software ecosystem is deeply entrenched, and its data centre revenue dwarfs every challenger combined. FuriosaAI is not going to displace Nvidia in training, and it does not claim to. The comparison that matters is narrower and more interesting: in inference, where the job is running models efficiently at scale, is there room for specialist chips that beat a general-purpose GPU on cost per token, power per token, or both?

FuriosaAI is one of several companies answering yes. Groq, whose $650 million round Infinitum co-led, builds chips optimised for extremely fast language model inference. Cerebras builds wafer-scale chips for both training and inference. A growing set of cloud providers, including Google, Amazon and Microsoft, design their own inference chips for internal use, which is a related but different model: they are not selling chips, they are lowering their own costs. And in the UK context, the most relevant comparison may be the one we covered in our article on Google’s TorchTPU challenge to Nvidia, where a hyperscaler is trying to break Nvidia’s grip by building its own silicon and software stack.

What distinguishes FuriosaAI within that field is a combination of three things. First, it is independent and merchant: it sells chips to anyone, rather than building them for a single cloud’s internal use. Second, it has reached volume production, which many challengers have not. Third, it has a clear geographic and policy story: a South Korean national champion in a sector where governments increasingly treat chip supply as a strategic asset. South Korea’s National Growth Fund putting $280 million into the round is not just a financial decision; it is an industrial policy decision, and it mirrors similar moves in the US, the EU and elsewhere to diversify AI hardware supply beyond a single dominant vendor.

That diversification matters for small businesses even if they never hear the name FuriosaAI. A market with one dominant supplier can charge what it likes, allocate scarce supply to its largest customers first, and set the terms on which software is built. A market with several credible suppliers, even if one remains largest, has to compete on price, on power efficiency, and on developer experience. The history of computing suggests that is how costs fall: not through a single breakthrough, but through sustained competition in a part of the stack that was previously taken for granted.

There are also limits to the comparison. Nvidia’s advantage is not just hardware; it is software, developer familiarity, and the sheer breadth of models and tools optimised for its platform. FuriosaAI must convince developers that its software stack is good enough, its compiler reliable enough, and its roadmap credible enough to justify supporting a second platform. Its collaboration with Broadcom on a third-generation accelerator for hyperscale infrastructure is relevant here: Broadcom brings deep experience in building chips at hyperscale volumes, and a credible third-generation roadmap helps customers commit to the second generation without fearing they will be stranded.

What it means for UK small businesses

As with Casap, the direct purchase question is the wrong starting point. UK small businesses will not buy RNGD cards, and they should not try to. The effects will arrive through the services they already use, and through the price and availability of AI features in those services. Here is the chain, step by step.

First, cheaper inference means cheaper AI features. Every time you use an AI writing assistant, a chatbot, a transcription tool, a search summariser or an image generator, someone is paying for inference. If inference hardware gets more efficient, the provider’s margins improve, and competition forces some of that improvement through to customers, either as lower prices, higher usage limits, or better models at the same tier. For a small business watching its software subscriptions, that is a meaningful, if slow-moving, benefit. It is also the reason the question of what AI costs, which we explored in our guide to how much AI costs UK small businesses in 2026, cannot be answered once and left alone: the cost base is moving, and it is moving downwards in inference even as frontier training costs rise.

Second, power efficiency matters for sustainability reporting and for energy costs. UK businesses face rising energy prices and growing pressure, from customers and from regulation, to report and reduce their carbon footprint. AI’s energy use is increasingly part of that conversation, and it is an awkward part, because a business rarely knows how much energy its AI features consume. Chips that deliver more work per watt, and data centres that can serve more customers from the same power budget, help on both fronts, even if the effect is invisible to the end user. A small business that chooses a provider partly on its efficiency story is making a bet that will look wiser, not less wise, as energy and reporting pressures grow.

Third, supply diversity reduces a quiet risk. Small businesses are at the end of the supply chain for AI hardware. When GPUs are scarce, large cloud providers and frontier labs get served first, and smaller customers face higher prices, longer waits, or throttled features. More suppliers, more production capacity, and more geographic diversity in chip manufacturing all reduce that risk. FuriosaAI’s plan to more than double RNGD production in 2027, and to build a third-generation chip with Broadcom, is a small contribution to that diversity, but in a market this concentrated, small contributions compound.

Fourth, there is a lesson about where to place bets. Small businesses are often advised to buy the market leader and not worry about the underlying technology. That advice is usually sound, but it is worth understanding where the leader’s advantage is structural and where it is contestable. In AI, training remains highly concentrated, but inference is more contestable, because the job is narrower and the performance criteria are clearer: cost per token, latency, and power. Businesses that understand that distinction will be better placed to judge vendor claims, to avoid lock-in to a single provider’s pricing, and to recognise when a cheaper alternative has become good enough for their use case.

Fifth, a practical implication for tool selection. When evaluating an AI tool, small businesses tend to focus on features and price, and both matter. But it is also worth asking, at least for tools where AI is a major cost driver, how the provider manages inference costs, whether it uses specialist hardware, and whether its pricing reflects efficiency gains over time. A provider that passes efficiency gains through to customers is a better long-term partner than one that pockets them while keeping prices fixed. You do not need to know the name of the chip inside the data centre to ask that question, but knowing that chips like RNGD exist, and that competition in inference is intensifying, makes the question sharper.

Honest limits / what to watch

As with any large funding round, the announcement should not be mistaken for a verdict. Several limits deserve attention.

Scale is the first. FuriosaAI plans to produce over 40,000 RNGD accelerators in 2027. That is a significant ramp for the company, but it is a fraction of the AI accelerator market, and it is not yet clear how much of that production is already committed to named customers versus how much must still be sold. The company’s deployments with LG, Upstage and Samsung SDS are encouraging, but they are not, on their own, proof of broad market adoption. The next year will show whether RNGD moves from a set of promising deployments to a product that a wide range of operators choose routinely.

Software is the second limit, and arguably the more important one. Hardware benchmarks are necessary but not sufficient. Developers need compilers, libraries, debugging tools and documentation that work reliably, and they need confidence that the platform will support the models they will want to run next year, not just the models that run well today. FuriosaAI’s support for Hugging Face, PyTorch and vLLM is a good start, but the test will be in the long tail: the unusual model, the new architecture, the edge case that a mature ecosystem handles without thinking. Nvidia’s ecosystem advantage was built over more than a decade, and it will not be overturned by a single generation of hardware, however efficient.

The third limit is geopolitical and supply-chain risk. FuriosaAI is a South Korean company operating in a sector where export controls, trade tensions and industrial policy all shape who can sell what to whom. Its US, Portuguese and Singaporean offices help, but they do not remove the risk that a future policy change affects its ability to serve certain markets or to source certain components. For UK readers, the relevant point is not to pick a side in that debate, but to recognise that hardware diversity is valuable precisely because it reduces dependence on any single country, company or supply chain.

Fourth, the valuation question. At about $2.9 billion, FuriosaAI is valued at a level that assumes substantial future growth, not just current revenue. That is normal for a chip company at this stage, and the investor list suggests sophisticated backers are willing to make that bet, but it also means the company must execute on production, software and sales at a pace that justifies the price. Chip companies have a long history of promising ramps that slip, and manufacturing at volume is unforgiving: yields, packaging, testing and delivery all have to work, not just the silicon design.

Finally, there is the question of what this means for AI’s environmental footprint overall. More efficient chips are good, but they can also lead to more AI being used, a rebound effect that economists have long observed with efficiency gains. If inference gets cheaper, people will run more inference, and total energy use may still rise even as energy per token falls. FuriosaAI’s sustainability claim is therefore best read as a claim about efficiency, not about total impact, and UK businesses that care about their footprint should look at total usage and total energy, not just efficiency per query.

Key facts

Timeline

• 2017: FuriosaAI is founded in Seoul, and DSC Investment, NAVER and KDB Capital back it at seed stage.
• January 2026: RNGD, the second-generation inference accelerator, enters mass production.
• January 2026: FuriosaAI begins fundraising for its Series D with a target of 450 billion to 750 billion won.
• July 2026: Samsung SDS launches a commercial NPU-as-a-service offering on RNGD.
• 28 May 2026 (reported): South Korea’s National Growth Fund management deliberation committee approves its investment in the round.
• 29 October 2026: FuriosaAI announces the completed $620 million Series D at about a $2.9 billion valuation.
• 2026–2027: Planned production rises from 20,000 RNGD accelerators in 2026 to more than 40,000 in 2027, alongside third-generation development.

What to watch next

Three signals will show whether this round delivers on its promise. The first is production: does FuriosaAI hit its 2026 and 2027 unit targets, and does it announce new customers beyond LG, Upstage and Samsung SDS? The second is software: do developers report that RNGD is easy to adopt, well documented, and reliable across a widening range of models? The third is the third-generation chip: does the Broadcom collaboration produce a hyperscale product on schedule, and does it attract the large cloud and enterprise customers that would turn FuriosaAI from a challenger into a fixture?

For UK small businesses, the takeaway is to watch inference costs, not just model launches. The next model announcement will grab headlines, but the quieter story, chips that run models more cheaply and efficiently, is what determines whether those models become affordable tools or remain expensive novelties. FuriosaAI’s $620 million is a bet that the quieter story is about to get louder.

Sources: FuriosaAI’s Series D announcement via Business Wire / Morningstar, 29 October 2026 (Korea time); Seoul Economic Daily, MoneyToday and Maeil Business Newspaper coverage of the round; Bloomberg reporting cited in AI Affairs coverage. Dollar conversions from won figures vary by exchange rate used; this article uses the company’s dollar figures as primary.

FAQs

What is FuriosaAI?

FuriosaAI is a South Korean semiconductor company founded in 2017 that designs AI accelerators for data centre inference, the work of running trained AI models to answer questions and complete tasks. Its second-generation chip, RNGD, entered mass production in January 2026.

How much did FuriosaAI raise, and at what valuation?

FuriosaAI raised $620 million (820 billion won) in a Series D round, bringing its post-money valuation to approximately $2.9 billion. South Korea’s National Growth Fund was the largest investor, and international investors contributed $110 million in total.

What is RNGD, and why does it matter?

RNGD is FuriosaAI’s second-generation AI inference accelerator, built on its Tensor Contraction Processor architecture. It is designed to deliver high throughput for running AI models while using less power than a comparable GPU setup, which matters because inference, not training, is where most AI cost accumulates over time.

What will FuriosaAI do with the new funding?

The company will expand RNGD production, strengthen its software ecosystem, and accelerate development of its third-generation accelerator, including work with Broadcom on a hyperscale chip. It expects to produce 20,000 RNGD accelerators in 2026 and more than 40,000 in 2027.

What does FuriosaAI’s funding mean for UK small businesses?

UK small businesses will not buy RNGD chips directly, but more competition in inference hardware should, over time, lower the cost of running AI features in the tools they use, from customer service to search, document processing and video. Cheaper inference is what turns AI demos into affordable everyday features.

Is FuriosaAI competing with Nvidia?

Yes, in inference, but not by trying to copy Nvidia. FuriosaAI focuses on a specific job, running trained models efficiently, using a different chip architecture and a power-efficient design, while Nvidia remains dominant across training and the broader AI computing market.

CTA: Get a Free AI Consultation