From Sharing a CES Stage to Becoming AMD's Chief Scientist: Nine Months After Fei-Fei Li's Appearance, What Changed at AMD and What Comes Next

On January 5, 2026, at CES in Las Vegas, Lisa Su invited Fei-Fei Li on stage during AMD's opening keynote. Li said at the time that some of World Labs' models were already running on AMD Instinct accelerators, and that the team had achieved performance gains using Instinct hardware and the ROCm software stack.

Nine months later, that shared stage turned into an $8.2 billion all-stock acquisition, with Li becoming AMD's chief scientist. Over those nine months, AMD went from an "AI chip chaser" to a company with a market cap above one trillion dollars. This article answers two questions: what exactly changed at AMD during this period, and what are the trends ahead?

Nine-Month Timeline

What Actually Changed at AMD Over These Nine Months

1. The data center went from a "side story" to the "main story." In the second quarter, AMD reported revenue of $11.54 billion, up 50% year over year, with a non-GAAP gross margin of 56%; data center revenue was $6.7 billion, up 107% year over year, accounting for 58% of total revenue. Third-quarter guidance is about $13 billion (±$300 million), up roughly 41% year over year. Lisa Su said revenue growth is expected to be "substantially above" the earlier target of more than 35%, and earnings per share will "significantly exceed" the previous $20 target.

2. From selling chips to selling racks. Helios is a full-rack system that bundles GPUs, CPUs, networking and software, with a price tag in the millions of dollars. That means AMD's competitor shifted from "another chip" to Nvidia's entire system, and it also means AMD's revenue mix, margin structure and delivery capabilities are all being rewritten.

3. A qualitative shift in the customer list. OpenAI, Anthropic, Microsoft, Meta and Oracle have all appeared on Helios or MI450 deployment lists. A year ago, AMD was mainly a "fallback option" in large-model AI labs; now it is an infrastructure supplier written into multi-year contracts.

4. Stepped-up investment in software and ecosystem. Reports say AMD's partnership with Anthropic includes joint optimization for ROCm; at MLPerf, software alone delivered a 70% per-cycle performance gain, suggesting ROCm's catch-up curve is getting steeper.

5. A faster M&A cadence. Taalas (inference) was followed by World Labs (model research), with AMD filling in inference hardware while extending upstream toward the "model side."

Five Trends Ahead

Trend one: Model research is pulled into the chip design loop. Fei-Fei Li reports directly to Lisa Su, a clear organizational signal: AMD wants "what the next-generation model needs" to be an input to the hardware roadmap rather than something to adapt to after the fact. The persistent scene state of world models is naturally more memory-bandwidth hungry, and we expect the generations after MI500 to lean more visibly toward such workloads in HBM capacity and interconnect (this is our inference; AMD has not commented officially).

Trend two: From "text" to the "physical world." AMD's official framing is to design AI hardware, software and systems for "inference, robotics, simulation and physical AI." AMD has the embedded and adaptive computing product lines that came with Xilinx, which have a natural place in robotics and at the edge—this is our judgment based on its product lines; specific product plans have not yet been announced.

Trend three: "Openness" becomes the differentiator. An "end-to-end open AI ecosystem" has to be delivered simultaneously across ROCm, open models and the platform. Nvidia's CUDA is a closed-source but extremely mature moat, leaving AMD only the differentiated path of openness plus low cost.

Trend four: Customer concentration and the risk of "investing for orders" are rising together. OpenAI's warrants and the roughly $5 billion investment commitment to Anthropic are essentially AMD using its own capital to pry loose orders from major customers. The bigger the orders and the deeper the ties, the greater AMD's dependence on a handful of AI labs; if these customers' fundraising pace slows, supply chain and inventory pressure will amplify in step. This is the most worrisome cyclical structure in the AI infrastructure industry today—it is not unique to AMD, but AMD is now deeply caught up in it.

Trend five: Competition with Nvidia becomes a "battle of cadence." Nvidia's Vera Rubin is already in full production, with quarterly data center revenue of about $89 billion, more than ten times AMD's. AMD's window lies in whether Helios can ramp on schedule, whether MI455X can deliver performance in real customer clusters, and whether ROCm can become "plug and play" across more workloads.

Milestones to Watch Next

1. Before the end of 2026: The World Labs deal closes, with disclosure of the specific share count, lock-up period and retention terms.
2. Second half of 2026: Helios shipment ramp; progress on OpenAI's first 1GW MI450 deployment.
3. Third-quarter 2026 earnings: Whether it can deliver on guidance of about $13 billion and a 56% gross margin, and whether data center's share continues to rise.
4. First quarter of 2027: Anthropic's first 1GW deployment.
5. The next MLPerf round: Whether it can hold its lead in video generation.
6. Stability of the World Labs team: Whether core researchers (such as Justin Johnson and Ben Mildenhall) stay on.

Our Take

Over nine months, AMD accomplished three things: it turned its AI roadmap (Helios, MI450/MI500) into deliverable products, locked major customers (OpenAI, Anthropic) into multi-year contracts, and bought "model research" into the company. From a single shared stage at CES to a signed acquisition agreement, this shows it was not an impulse but a simultaneous bet by AMD across all three layers: hardware, software and models. The risks are equally clear: the stock is already at a record high, dependence on a small number of customers is deepening, and the real contribution of the acquired team will not be verifiable for years. AMD's task right now is not to tell an even bigger story, but to get Helios shipping on schedule.

This article is based on public reporting and company disclosures and does not constitute investment advice.

Sources

• Tom's Hardware: AMD acquires AI legend Fei-Fei Li's World Labs
• AMD IR: Second Quarter 2026 Financial Results
• TradingKey: AMD Q2 2026 Earnings Analysis
• TradingKey: Advancing AI 2026 Roundup
• AMD Newsroom: CES 2026 Keynote
• Futurum: MLPerf and World Models Analysis