AMD, Furiosa AI and MangoBoost set up a heterogeneous AI infrastructure cluster in Seoul: the alternative to Nvidia advances

AMD, FuriosaAI and MangoBoost are building a heterogeneous AI infrastructure cluster in Seoul: the Nvidia alternative moves forward

🔎 The GPU monopoly starts to crack

On November 4, 2026, during the AMD Advancing AI World Tour in Seoul, AMD signed a memorandum of understanding (MOU) with FuriosaAI and MangoBoost, two Korean startups specializing respectively in inference NPUs and DPUs. The stated goal: to assemble a reference cluster in Seoul by spring 2027, combining CPUs, GPUs, NPUs and DPUs within a single architecture. The news was confirmed by Chosunbiz and then picked up by DigiTimes and Money Today.

Why now? Because the AI economy has shifted toward inference, and inference punishes homogeneous architectures. The Stanford AI Index 2025 documents a 280-fold drop in inference costs over two years: margins are shifting toward whoever can deliver tokens at the lowest cost per watt. That is exactly the playing field of a heterogeneous cluster.

And because supply chain concentration has become a strategic risk that buyers themselves now acknowledge. When Google agrees to pay SpaceX $920 million per month for 110,000 Nvidia GPUs, the message is clear: demand exceeds supply, and every serious buyer is looking for a plan B. Seoul has just sketched one out, in very concrete terms.

The Essentials

• AMD, FuriosaAI and MangoBoost sign an MOU on November 4, 2026: a heterogeneous reference cluster (CPU + GPU + NPU + DPU) in Seoul, operated by MangoBoost, targeted for spring 2027 (Chosunbiz).
• Target: the cost of inference, by assigning each task to the most efficient processor, not the most powerful one.
• Context: Nvidia holds roughly 90% of AI accelerators (analyst estimates, 2025), US hyperscalers have announced over $380 billion in capex for 2025, and DRAM has remained under strain since late 2025 (TrendForce).
• For Korea: an integration of its national champions into a global platform, with separate discussions between AMD and DEEPX on robotics (DigiTimes, November 5, 2026).
• For businesses: the era of "Nvidia-only" purchasing warrants a plan B, measured in tokens/second/watt.

Recommended Tools

What AMD Actually Announced in Seoul

An MOU, not a contract: and a reference cluster in spring 2027, not a volume product. According to Chosunbiz, the agreement signed on November 4, 2026 aims to test AI systems with mixed chips and to reduce inference costs. The cluster will be operated by MangoBoost in Seoul, a senior vice president at AMD specified.

Concretely, the project brings together three complementary building blocks. AMD provides the server computing power (Instinct GPUs, EPYC CPUs) and the ROCm software layer. FuriosaAI provides its NPUs, optimized for low-power inference. MangoBoost provides its DPUs, which offload networking and storage — the "east-west" traffic that chokes agent clusters.

AMD is also holding separate discussions with DEEPX, the other Korean NPU champion, about low-power chips intended for robotics (DigiTimes, November 5, 2026). The signal is clear: AMD is not looking for one more customer, it is building a national ecosystem.

The Three Partners at a Glance

This trio is a strategic choice by AMD, and I find it clever. Rather than a frontal war against CUDA, AMD positions itself as the orchestrator of an open ecosystem where each silicon plays its part. It's less spectacular than a "Nvidia killer," but it's more credible.

What This MOU Is Not

It is neither an exclusivity, nor a production line, nor a catalog product. A demonstration MOU tests an architecture, it does not yet sell it. Volumes, if they follow, will come after the public benchmarks — and after the integrators have proven that managing a mixed GPU+NPU+DPU fleet doesn't require an army of engineers.

Heterogeneous: the end of "one GPU for everything"

Because in production, the bulk of the AI workload is inference, and inference doesn't need the most expensive GPU in the world for every token. This is the most firmly established analyst consensus of 2025: training concentrates capital, inference concentrates operating costs.

The numbers set the tone. The Stanford AI Index 2025 measures a roughly 280x drop in the inference cost of a GPT-3.5-level model between late 2022 and late 2024. This price pressure imposes an industrial logic: every task must run on the least expensive silicon capable of handling it.

This is where heterogeneity takes on its full meaning. NPUs like FuriosaAI's RNGD efficiently handle stable inference workloads for a few hundred watts, versus the kilowatt of a high-end GPU. MangoBoost's DPUs absorb network traffic and storage access, freeing up the GPUs for compute. Every watt spent serves useful computation, not overhead.

Disaggregation, the common thread with Helios

This logic aligns exactly with what we documented in our analysis AMD Helios + Cerebras: 5x tokens/sec/watt on models >1T parameters: decoupling prefilling, decoding, and memory to maximize throughput per watt. The Seoul cluster applies the same philosophy at the scale of a country, with Korean silicon on top of that.

Worth noting, to be fair: Nvidia has itself been heterogeneous for years, with its Grace CPUs and BlueField DPUs. The difference is that Nvidia sells a 100% Nvidia stack. The Seoul cluster is betting on the opposite: an open, composable stack, where AMD doesn't hold every piece.

Nvidia under simultaneous pressure: the numbers

No, this cluster won't dethrone Nvidia anytime soon; yes, it attacks where competition becomes possible: the economics of inference and access to supply.

The fundamentals remain overwhelming. Nvidia posted $57 billion in revenue in the third quarter of its fiscal year 2026, including $51.2 billion from data centers (results published in November 2025). Its market share in AI accelerators hovers around 90% according to analysts. The combined capex of the four American hyperscalers exceeded $380 billion for 2025 alone, according to their quarterly reports.

But the crack lies elsewhere: in dependency. When a customer the size of Google has to rent 110,000 GPUs via SpaceX for $920 million per month — our analysis here —, it's a sign that even the giants can't get what they want, when they want it. This tension mechanically creates demand for any credible alternative.

The regulator is getting involved too. The US DOJ is already looking into Nvidia's acquisition of Hugging Face for $1.29 billion. A dominant position this visible attracts antitrust attention, and every investigation strengthens buyers' appetite for a second source. The timing is telling: in 2026, Gartner projected nearly $2 trillion in global AI spending (September 2025 forecast). Nobody wants to concentrate all of that on a single supplier.

My take is simple: Nvidia's monopoly is strongest where it is integrated — training and CUDA — and most fragile where buyers optimize cost per token. The Seoul cluster targets only that second front. It's the right choice, and that's why it's moving forward.

Korea and Europe get organized: the supply chain is regionalizing

The real lesson of the announcement: the AI supply chain is regionalizing, and Seoul is assembling precisely the kind of national stack that tech sovereignties have been demanding for two years.

Korea starts with real assets. SK hynix dominates the HBM market, that high-bandwidth memory that every AI accelerator depends on. Samsung carries weight in advanced foundry. FuriosaAI, founded in 2017, already supplies players like Naver with its Warboy chips, and its RNGD, announced in 2024, delivers 512 TOPS in INT8 for roughly 180 W. MangoBoost, founded in 2021 by professor Jangwoo Kim (POSTECH), is establishing itself as the reference for DPUs for full-stack "agentic" AI infrastructure, as the company's website claims.

The detail that says it all: FuriosaAI turned down an acquisition offer of around $800 million from Meta in 2025 (Reuters, March 2025) to stay independent. Twelve months later, that independence has earned it a central role in a cluster sponsored by AMD. Founders who refuse buyouts are sometimes offered something better than the buyout: a market.

Europe is following the same path, each at its own scale. The Vera Rubin NVL72 racks assembled in France by Bull and Foxconn show that even Nvidia is willing to regionalize its production under political pressure. With AMD allying itself with Seoul and Nvidia manufacturing in Europe, the geography of AI is being reshaped before our eyes.

My honest opinion: AI sovereignty is not decreed through press releases, it is built through industrial contracts and public benchmarks. Seoul just signed one. All that remains is to honor it.

The real obstacle: software, not silicon

If the Seoul cluster disappoints, it won't be for lack of TOPS, but for lack of software maturity. It's the one thing press releases never mention, and it's what will decide everything.

CUDA remains Nvidia's defensive moat: eighteen years of tooling, millions of developers, a library ecosystem impossible to replicate in two years. ROCm has come a long way — PyTorch and vLLM run on recent Instinct cards — but porting any exotic model is still a project, not a checkbox.

On top of that comes the fragmentation inherent to heterogeneity. A mixed GPU+NPU+DPU cluster requires a scheduler capable of routing each task to the right processor, proprietary NPU SDKs to master, and DPU drivers to maintain. Every building block added brings another failure surface. That's the price of efficiency, and MangoBoost will have to pay it with its demonstration cluster.

Three signals to watch in 2027

1. Public benchmarks of the Seoul cluster (spring 2027): tokens/sec/watt measured, not promised.
2. Helios deliveries to OpenAI: the October 2025 agreement calls for 6 gigawatts of Instinct systems across multiple generations, including 1 GW of MI450 as early as the second half of 2026. If AMD keeps to that schedule, Seoul becomes credible.
3. Native support for NPU/DPU offload in inference frameworks (vLLM, SGLang): the day heterogeneity becomes a configuration flag rather than an engineering project.

CTO action plan: prepare without breaking everything

Build a dual-sourcing plan now: Nvidia remains your foundation for training, and you launch a measurable inference pilot on alternatives before price pressure forces it on you.

Concretely, in order:

1. Inventory your workloads: token volume, target latency, current cost per million tokens.
2. Choose self-hostable models to test without depending on APIs. The table below ranks the best options.
3. Test inference on Instinct via a cloud offering MI355X: the ROCm stack has matured on standard models.
4. Measure what matters: tokens/second/watt, cost per million tokens, p99 latency. Not TOPS.
5. Negotiate a multi-vendor clause into your next infrastructure contracts: the mere fact of asking for it changes your prices.
6. Set up your staging elsewhere: a KVM VPS from Hostinger (from €4.99/month, November 2026, verify on hostinger.com) is enough to host the gateway for your self-hosted agent. Leave the GPUs to the cloud.

*Agentic ranking, public data 2026.

The gap between the best API model (98.2) and the best self-hostable one (88.1) is the price of your independence today. Ten points, on agentic benchmarks, is significant for critical use cases — but largely acceptable for support, enterprise RAG, or bulk processing. And this gap narrows with every generation, while self-hosted inference costs follow the market's downward curve.

This is exactly why heterogeneous computing interests the Koreans: open models like Kimi K2.6 or GLM-5 are the ideal workloads for NPUs and cost-optimized infrastructure. One doesn't go without the other.

❌ Common Mistakes

Mistake 1: Confusing an MOU with a shipped product

What goes wrong: many commentators treat every AI memorandum of understanding as a revolution delivered the very next day. Yet the Seoul cluster is a demonstration scheduled for spring 2027, not a commercial offering. The solution: plan on Nvidia as your backbone through at least 2027, and treat the alternatives as measured pilots, not migrations.

Mistake 2: Buying accelerators based on TOPS

What goes wrong: TOPS don't predict tokens/second. Memory bandwidth, memory capacity, and above all software maturity dominate real-world performance. A poorly utilized 512-TOPS NPU will underperform a GPU with half its theoretical specs. The solution: benchmark against your actual workload, measuring tokens/sec/watt, p99 latency, and cost per million tokens.

Mistake 3: Waiting for "the Nvidia killer"

What goes wrong: looking for a single product that replaces Nvidia everywhere is a category error. The market is fragmenting by workload, not around a panacea: Nvidia keeps training and extreme performance, while the alternatives win cost-effective inference and the edge. The solution: mix vendors by use case, and make your trade-offs on TCO, not on press releases.

❓ Frequently Asked Questions

What is a "heterogeneous" AI cluster?

It's a cluster combining several types of processors — CPUs, GPUs, NPUs, DPUs — where each handles the tasks it is most efficient at. Instead of a single type of GPU doing everything, network traffic goes through DPUs, routine inference through NPUs, and heavy computation through GPUs. This is the architecture that AMD, FuriosaAI and MangoBoost are assembling in Seoul.

When will the Seoul cluster be operational?

The reference cluster, operated by MangoBoost in Seoul, is planned for spring 2027, according to an AMD senior vice president quoted by Chosunbiz (November 4, 2026). The MOU was signed on November 4, 2026 during the AMD Advancing AI World Tour. Note: this is a demonstration, not a commercial offering at scale.

Is this the end of Nvidia's dominance?

No, not in the immediate term. Nvidia holds roughly 90% of the AI accelerator market and $51.2 billion in data center revenue in Q3 of its fiscal 2026. But economical inference is a playing field where CUDA carries less weight, and every alternative that proves its energy efficiency gains share there. Dominance will erode first on tokens, not on training.

What does this change for a European company?

Nothing operational before 2027, but a lot strategically. The manufacturing of Vera Rubin NVL72 racks by Bull and Foxconn in France, combined with the rise of Korean alternatives, strengthens your negotiating power. Start requiring multi-vendor options in your infrastructure contracts now, and test non-Nvidia inference pilots.

Which open models can you self-host today?

Moonshot AI's Kimi K2.6 (agentic score of 88.1 self-hosted) and Z.AI's GLM-5 (Reasoning) (82) are the best options at the moment. Expect a real engineering workload for hardware optimization, but these models already allow you to decouple your agents from API pricing — and they are the natural first customers of heterogeneous clusters.

✅ Conclusion

Seoul's heterogeneous cluster doesn't kill Nvidia, but it materializes what every AI infrastructure buyer has been waiting for: an organized, financeable, and measurable alternative, where CPU, GPU, NPU, and DPU work together at the best cost per token. To go further, read our analysis of AMD Helios + Cerebras and disaggregated inference: it's the same revolution, at rack scale.