One of the longest-running rivalries in consumer computing surfaced on Saturday, October 24, on a stage no one had put in writing. In a conversation with Jimmy Cheng, president of CASPA, the association of Chinese and American semiconductor professionals, Intel CEO Lip-Bu Tan said he has spoken with Lisa Su, his counterpart at AMD, about areas where the two companies could cooperate. He did not give a single hint of what they are, and neither company has announced any agreement.
The conversation does not emerge in a vacuum. Since October 15, 2024, Intel and AMD have maintained the x86 Ecosystem Advisory Group, a forum with Broadcom, Dell, Google, HPE, Lenovo, Meta, Microsoft, Oracle and Red Hat so that software does not break when switching processors, plus the endorsements of Linus Torvalds and Tim Sweeney. On October 20 that forum bore its first visible fruit: AMD put ZenDNN into oneDNN, the library with which PyTorch accelerates on CPU, and the two companies aligned their approaches under the umbrella of the UXL Foundation.
What is at stake is not a merger or an industrial alliance, but who foots the bill for keeping x86 alive while Arm gains ground in laptops and data centers. For anyone buying a processor, the useful part is more modest than the headline suggests: software that already runs is starting to receive common optimizations, and the bill does not change immediately. Tan has a plan for it to change later on, backed by custom chips and co-packaged optics.
One talk, one sentence and no figures
The recording of the October 24 talk leaves Tan discussing rebuilding the CPU business long before reaching AMD. The conversation with Lisa Su appears as a loose aside, with no specific areas, no deadlines and not a single figure to back it up.
His most forceful line in the entire talk was an internal promise: the company must regain leadership in CPU, and he is determined to achieve it. The diagnosis he made of his own company proved harsher than any comparison with the competition, because he admitted that Intel never wanted to adapt a chip to what each customer asked for.
Now it adapts the product to the order, he says, and he intends to double and triple down on that bet. Customization stops being a laboratory extra and comes to be treated as a business line with revenue of its own, something the company had not set out with this level of detail.
Underneath that shift are four bets that Tan placed on the same level as the usual CPUs, and that until now lived on the margins of the catalog. They are the lines that should sustain growth when desktop processor share can go no further, because that market grows little and competes with Arm on price.
- Custom chips (ASIC) for very specific workloads of a single customer.
- Co-packaged optics, which brings fiber to the package itself to move data with less energy.
- Physical AI, the kind that goes inside robots and machines operating in the real world.
- Advanced packaging, the technique of joining several pieces of silicon into a single product.
It wants two or three of those lines to contribute around 10% of revenue each, and there the familiar problem surfaces: these are percentages of a business that does not yet exist at that scale. The timetable is missing, and so is the customer who signs first.
| Part of the plan |
Status as of October 26 |
| Intel AMD Lisa Su conversation |
Confirmed by Tan on Saturday the 24th, with areas not detailed |
| New agreement between the two companies |
None announced |
| x86 Ecosystem Advisory Group |
Active since October 15, 2024 |
| ZenDNN inside oneDNN |
Announced on October 20, with no dated delivery in PyTorch |
| 14A with PDK 1.0 for volume |
Expected for early 2027 |
x86 has shared a library for a week
The October 20 announcement is not a statement of intent. The two companies have aligned oneDNN as the default library for x86 acceleration within PyTorch, and AMD has joined as a member of the steering committee of the UXL Foundation, which governs the project from a neutral position.
The unification of PyTorch optimizations in oneDNN closed a duplication we had been carrying for years. Until then, an EPYC went through zentorch, AMD's own plugin, while a Xeon used Intel's path. Two code stacks, two teams and the same underlying problem, because maintaining duplicated kernels costs money and delays every version of the framework.
The agreement goes beyond moving files from one repository to another. Both companies are also working on AI Compute Extensions, an extension of the x86 repertoire aimed at matrix multiplication, the operation that dominates neural network inference. It will reach PyTorch through future versions of oneDNN, with no public arrival date.
The factory is another matter
That is where the cooperation stops. Tan said that the 18A process is already in volume production and that 14A is on track: version 0.5 of the design kit has been delivered, 0.9 comes next and 1.0 will arrive for volume in early 2027.
Two potential customers are evaluating 14A test wafers, with firm decisions expected between the second half of this year and the first half of next year. AMD is not on that list and has no reason to be, because the company does not manufacture: it buys wafers from TSMC for all of its 7-nanometer or smaller microprocessors and GPUs.
Competing fab against fab and cooperating on software are different businesses, and the conversation with Lisa Su only makes sense in the second of them. Tan described the foundry as a business where trust weighs as much as technology, because the customer hands over wafers on which its own revenue depends.
He recalled a customer who waited six or eight months for him to become CEO to present him with fifteen points where Intel had failed. As he resolved them one by one, the customer agreed to expand the relationship.
What the processor buyer notices
Little, and it is worth saying so. There is no version of PyTorch to announce with these optimizations inside, nor a delivery date: AMD's work on the common library is a plan with an open calendar. The useful check, when they come out, is to read the release notes before touching production.
The underlying pulse is the same as the requirements Nvidia set for RISC-V to host CUDA, with the RVA23 profile as the entry key: whoever controls the compiler and the libraries decides which hardware is worth buying. Intel and AMD have just admitted that neither of them can maintain its optimization stack alone.
Two rivals with the same library and different fabs
The headline question has a short answer: there is no alliance, there are conversations and a shared library that already works. Cooperation will come from where it is not seen, in compilers and kernels, and it will stay there as long as manufacturing for third parties remains the ground where Intel is staking its future.
For the buyer, the practical consequence is that the processor brand will weigh less in the code and more in the price and in the support. On that, yes, they can decide today, without waiting for the two companies to clarify themselves, and with the same information the two CEOs have: a common library, two separate fabs and no signed document.