Here is what I think investors misunderstand about AMD.
They look at Nvidia’s dominance and assume AMD must defeat Nvidia for the investment thesis to work.
It doesn’t.
AMD does not need to replace Nvidia. It needs to become the credible second platform that every hyperscaler, AI laboratory, and sovereign-computing project wants available.
And that opportunity could be enormous.
No One Wants a Single Supplier
Nvidia currently has the strongest AI platform in the world.
The company does not simply sell GPUs. It sells an integrated system of chips, networking, software, libraries, and developer tools that customers already know how to use.
That creates a tremendous competitive advantage.
But it also creates a problem for Nvidia’s customers.
Microsoft, Meta, OpenAI, Anthropic, and the other large buyers of AI infrastructure do not want the future of their businesses controlled by one supplier.
They want more capacity.
They want negotiating leverage.
They want control over their infrastructure roadmaps.
And they want an alternative if Nvidia cannot provide enough systems at the price, performance, or schedule they need.
That is where AMD becomes extremely important.
AMD’s opportunity is not simply to produce a faster GPU.
Its opportunity is to give the AI industry a second complete ecosystem.
AMD Is No Longer Just Selling Chips
A few years ago, the AMD AI thesis was relatively simple.
AMD would build an accelerator, sell it to cloud providers, and attempt to take a small amount of market share from Nvidia.
That is no longer the strategy.
AMD is now building the entire rack.
The company’s Helios rack-scale platform combines Instinct GPUs, EPYC server CPUs, Pensando networking, and ROCm software into one integrated AI system.
A full Helios rack includes 72 MI455X GPUs and 18 sixth-generation EPYC processors, connected through AMD’s networking architecture.
This matters because the unit of competition in AI is changing.
The market is no longer comparing one GPU with another GPU.
Customers are comparing complete systems.
How many tokens can the rack generate?
How much electricity does it consume?
How much memory does it provide?
How reliably can thousands of accelerators communicate?
How quickly can the system be installed, cooled, and placed into production?
That is the game Nvidia has been playing with its rack-scale systems.
Helios is AMD’s attempt to prove it can play that game too.
Memory Could Be AMD’s Opening
AI models require enormous amounts of fast memory.
That becomes especially important during inference, when the system must hold model weights, context, and the growing key-value cache created as users interact with the model.
The MI455X is designed with 432 gigabytes of HBM4 and up to 23.3 terabytes per second of memory bandwidth. Across a complete Helios rack, that adds up to approximately 31 terabytes of HBM4.
That is not a minor specification.
Additional memory can allow customers to run larger models, handle longer context windows, serve more simultaneous users, and reduce the number of systems required for memory-intensive workloads.
This is where AMD may be able to differentiate.
Not every AI customer needs the accelerator with the highest theoretical performance.
Some customers may care more about memory capacity, total system cost, availability, energy efficiency, or how many useful tokens the rack can produce per dollar.
AMD claims Helios can deliver up to 30% more inference tokens per dollar than a competing rack on a specific Kimi K2 Thinking workload.
That is an internal AMD estimate—not an independent verdict covering every model and workload.
But it tells us exactly where AMD intends to compete.
Not on hype.
On economics.
The Customers Are Becoming Difficult to Ignore
A product roadmap is interesting.
A product roadmap accompanied by gigawatt-scale customer commitments is considerably more interesting.
OpenAI and AMD announced an agreement covering up to six gigawatts of AMD GPUs across multiple product generations.
Meta announced plans to deploy up to six gigawatts, beginning with a one-gigawatt deployment using a custom MI450-based accelerator.
Anthropic agreed to deploy up to two gigawatts of MI450-series GPUs, with its first gigawatt expected to begin deployment during the first half of 2027.
Microsoft also plans to deploy Helios at scale through Azure, using the systems for frontier-model inference, Azure AI services, and customer workloads.
Those announcements do not guarantee that every proposed gigawatt will arrive on schedule or generate the revenue investors expect.
But they do answer one extremely important question.
The largest AI companies are taking AMD seriously.
These customers are not experimenting with a few accelerators in a laboratory.
They are aligning hardware, software, and infrastructure roadmaps around potential deployments measured in gigawatts.
That is a completely different level of validation.
EPYC Is the Part People Forget
The AMD AI story is not limited to Instinct GPUs.
Every accelerator still needs CPUs to orchestrate workloads, prepare data, manage storage, run control-plane operations, and keep the rest of the system moving.
AMD’s EPYC processors already have a major presence across cloud and enterprise data centers.
That gives AMD an advantage Nvidia did not originally have.
AMD can sell the CPU.
AMD can sell the GPU.
AMD can provide the networking through Pensando.
AMD can provide the software through ROCm.
And now AMD can help integrate everything into the rack.
The company is moving from selling individual components to capturing more of the value inside each AI deployment.
The financial results are beginning to reflect that shift.
AMD’s second-quarter 2026 data-center revenue reached $6.7 billion, increasing 107% year over year and representing approximately 58% of total company revenue.
That figure includes both EPYC processors and Instinct accelerators, so we should not pretend all $6.7 billion came from AI GPUs.
But the broader message is clear.
Data centers have become the center of AMD’s business.
ROCm Is Still the Entire Ballgame
Hardware specifications get attention.
Software determines whether customers stay.
Nvidia’s CUDA ecosystem has been developed, optimized, and adopted for years. Developers know it, companies have built around it, and enormous amounts of AI software already run on it.
That is Nvidia’s real moat.
AMD can build an outstanding accelerator and still lose if deploying models on it requires too much engineering work.
That is why ROCm may be the most important part of the AMD thesis.
AMD needs models to work immediately.
It needs PyTorch, JAX, Triton, vLLM, and the rest of the modern AI software stack to perform reliably.
It needs customers to move workloads between Nvidia and AMD without rebuilding everything from the ground up.
And it needs developers to view AMD as a normal production choice—not a science project requiring constant customization.
The OpenAI and Anthropic partnerships could be particularly valuable here because those companies are not merely purchasing hardware.
They are working with AMD to optimize real frontier-model workloads across the hardware and software stack.
That creates a feedback loop.
More large customers improve the software.
Better software attracts more customers.
More customers justify more optimization.
That is how an ecosystem begins to compound.
Open Could Become a Competitive Weapon
Nvidia’s greatest strength is control.
AMD is betting that openness can become a competing advantage.
Helios is based on open industry standards rather than an entirely proprietary rack architecture. ROCm is open source, and AMD is supporting technologies such as UALink and Ultra Ethernet.
The pitch is simple.
Customers should be able to choose components, customize systems, and avoid becoming permanently dependent on one vendor’s architecture.
This will not matter to every customer.
Some will happily accept a proprietary system if it provides the best performance and easiest deployment.
But the largest cloud providers design infrastructure on a scale where flexibility has enormous value.
They do not want to rent their future from a single supplier.
AMD gives them another option.
AMD Still Depends on the Same Physical World
AMD may offer an alternative to Nvidia, but it does not escape the supply chain.
Its accelerators still require advanced manufacturing.
They still need HBM.
They still need sophisticated packaging.
They still need optical networking, power, cooling, and enormous data-center construction.
The MI455X uses advanced CoWoS-L packaging and HBM4, which means AMD is competing for many of the same constrained resources Nvidia needs.
AMD has announced more than $10 billion in Taiwan ecosystem investments to expand partnerships and help scale the manufacturing and packaging required for Helios.
But an announcement is not a finished rack.
AMD still has to receive enough silicon and memory, assemble the systems, deliver them on schedule, and help customers place them into production.
Demand means very little if the company cannot ship.
What Could Go Wrong
The first risk is software.
ROCm has improved significantly, but “improved” is not the same as “equal to CUDA across every important workload.”
The second risk is execution.
Helios is AMD’s first serious attempt to compete as a complete rack-scale platform. Integrating GPUs, CPUs, networking, memory, cooling, and software at this scale is incredibly difficult.
The third risk is that announced commitments may not become revenue as quickly as investors expect.
“Up to six gigawatts” is not the same thing as six gigawatts already installed and generating income.
Deployments can be delayed.
Customer requirements can change.
Product roadmaps can slip.
And Nvidia will not stand still while AMD attempts to close the gap.
The fourth risk is economics.
Vendor benchmarks are useful, but they are highly dependent on the model, batch size, latency target, utilization, software version, and power assumptions.
AMD must prove its tokens-per-dollar advantage in real production environments—not only in carefully selected comparisons.
What to watch out for
Watch whether Helios shipments ramp on schedule during the second half of 2026.
Watch whether Microsoft, Meta, OpenAI, and Anthropic move from announcements to operating deployments.
Watch data-center revenue, operating margins, and whether AMD begins providing greater visibility into Instinct sales.
Watch HBM4 availability and advanced-packaging capacity.
Watch how quickly ROCm supports new frontier models and whether customers can deploy them without substantial custom engineering.
Most importantly, watch independent tokens-per-dollar performance under real production workloads.
Because AMD does not win when a slide says Helios is competitive.
AMD wins when customers deploy it, developers use it, and the systems generate enough economic value that buyers order more.
Final Thoughts
I do not think the AMD thesis depends on Nvidia failing.
The AI infrastructure market could grow dramatically while Nvidia remains the dominant company.
AMD only needs to capture a meaningful portion of that growth.
If the world builds multiple gigawatts of AI infrastructure every year, becoming the credible second full-stack platform could create an enormous business.
The CPUs are already proven.
The GPUs are improving.
The customer commitments are arriving.
The rack-scale platform is entering production.
Now AMD has to execute.
That is the entire thesis.
AMD does not need to become the next Nvidia.
It needs to become the alternative that the AI industry cannot afford to ignore.