You already have an account? Log In

DeepSeek V4, AMD's MI355X and the Race to Make AI Cheaper to Run

Posted on 04.08.2026

Two years ago, running a frontier-class large language model required a data centre's worth of NVIDIA H100s and a budget most Australian businesses could only dream about. Today, the picture looks very different. DeepSeek's latest 1.6-trillion parameter release is being benchmarked not just on NVIDIA's newest silicon, but on AMD and Huawei hardware too — and the cost curve for inference is bending down fast.

That shift matters well beyond Silicon Valley. It changes who gets to build with AI, how much it costs to serve customers, and whether the sovereign-AI ambitions being floated in Canberra and elsewhere are actually affordable. Here's what's really going on beneath the benchmark charts.

The DeepSeek V4 benchmark story

SemiAnalysis has been tracking DeepSeek V4 1.6T performance from Day 0 to Day 43 across a suddenly crowded field: Huawei's Ascend accelerators, NVIDIA's GB300 NVL72 systems, NVIDIA's B200 and AMD's MI355X. The takeaway isn't which chip "wins" on any given day — it's that a Chinese open-weights model at 1.6 trillion parameters is now being optimised, in near-real-time, across four different hardware ecosystems.

That's the accessibility story in a single sentence. When DeepSeek dropped V3 late last year, the shock was how cheaply it had been trained. V4 is showing us the second act: how quickly the wider ecosystem can pick up a new open model and make it run efficiently on whatever silicon a customer happens to have.

The Day-0-to-Day-43 framing matters because early performance numbers on new models are usually dire. Kernels aren't tuned, memory layouts are wrong, and inference stacks haven't caught up. The fact that multiple non-NVIDIA platforms are now closing that gap in weeks rather than quarters is the underreported part of the story.

AMD's slow chip away at the CUDA moat

NVIDIA's real product has never really been GPUs — it's been CUDA, the software layer that has locked developers in for the better part of two decades. SemiAnalysis's recent piece framed around AMD Advancing AI 2026 asked bluntly whether AMD can finally break that moat.

The honest answer is: not overnight, but the cracks are widening. AMD's MI355X is showing up in serious deployments. Zyphra, an AI research lab, has just launched a new open AI platform powered by 15 megawatts of MI355X GPUs, with a roadmap that pushes on to the next-generation MI450 and beyond. Fifteen megawatts is not a science project — that's a real production-scale cluster, the kind that only made commercial sense on NVIDIA silicon a year ago.

Two things are driving this. First, AMD's ROCm software stack has matured to the point where popular open models — DeepSeek V4 among them — can be ported without a full-time team of CUDA whisperers. Second, memory bandwidth and HBM capacity on parts like the MI355X are genuinely competitive for inference workloads, which is where the money actually is once a model has been trained.

Why inference economics are the whole game

Training a frontier model is a one-off spend. Serving it to millions of users, every day, forever, is where costs accumulate — and where hardware choice matters most. For an Australian SaaS company embedding an LLM into its product, the difference between 2 cents and 0.2 cents per thousand tokens is the difference between a viable feature and a P&L disaster.

Three forces are compressing those costs simultaneously:

Model efficiency: DeepSeek's mixture-of-experts architecture activates only a fraction of its 1.6T parameters per token, dramatically cutting the compute needed per query.
Hardware competition: With AMD, Huawei and custom silicon from hyperscalers now credible alternatives, NVIDIA can no longer price without regard to substitutes.
Software maturity: Inference engines like vLLM, SGLang and TensorRT-LLM have compressed the same model onto far less silicon than was possible even 12 months ago.

Stack those together and you get order-of-magnitude improvements in tokens-per-dollar within a single generation. That is the quiet revolution the benchmark charts are pointing at.

NVIDIA isn't standing still — it's changing shape

None of this means NVIDIA is in trouble. Wccftech has reported that NVIDIA is on track to become the world's leading CPU supplier, with its Vera CPU line projected to hit US$20 billion in revenue this year. Read that sentence twice: the company famous for GPUs is about to become the biggest CPU vendor on the planet, largely on the back of tightly integrated CPU+GPU platforms like GB300 NVL72.

That vertical integration is NVIDIA's answer to AMD's price-performance push. If you can't win purely on chip-versus-chip economics, you win on the rack — networking, memory coherence, software, and now the CPU itself all designed to work together. For customers, this creates a genuine choice: pay the premium for a tightly integrated NVIDIA rack that just works, or assemble a cheaper AMD-based system and invest in the engineering to make it sing.

What this means for Australia

For Australian buyers — whether that's a fintech in Sydney, a mining major in Perth running document-analysis models, or a university deploying research AI — the shift has three practical consequences.

Cloud pricing will keep falling. AWS, Azure and Google Cloud all now offer AMD Instinct instances alongside NVIDIA ones. As models like DeepSeek V4 get properly optimised for MI355X-class hardware, expect a real second price tier to emerge for inference workloads.

Open-weights models become the default. When DeepSeek, Llama and Qwen-class models can be self-hosted efficiently on multiple vendors' silicon, the argument for locking into a proprietary API becomes harder to sustain — particularly for organisations with data-sovereignty concerns.

Sovereign AI becomes plausible. A 15MW cluster like Zyphra's is not a hyperscaler-only proposition anymore. State governments, universities and larger enterprises can realistically consider standing up their own inference capacity, especially as MI450 and future generations arrive.

The catch: efficiency doesn't equal accessibility

It's worth a note of caution. Cheaper per-token inference doesn't automatically democratise AI. The organisations best placed to exploit these gains are the ones already staffed with ML engineers who can tune ROCm kernels, wrangle vLLM configs and evaluate model quality across hardware. For a small business, the practical accessibility gain still comes through APIs from OpenAI, Anthropic, Google or their Australian resellers.

But the wholesale cost of intelligence — the underlying price of running a token through a frontier model — is dropping in a way that will eventually reach every layer of the stack. DeepSeek V4 running credibly on four different hardware platforms within 43 days of release is the clearest signal yet that we've entered a genuinely competitive era of AI infrastructure.

The chip wars, in other words, are finally starting to benefit the people paying the bills.