There's a version of the AI story that gets told a lot right now. Chips. Data centers. Billions of dollars. Nvidia. Everyone has heard it.
What doesn't get told as often is what actually happens when all that infrastructure starts getting used. That's the part Lisa Su was talking about on July 24 in San Francisco, and it's the part that matters most for what comes next.
Su was on stage at AMD's Advancing AI 2026 conference at the Moscone Center. She's not someone who reaches for dramatic language. So when she said AI has hit an inflection point where people are doing genuinely useful, meaningful work with it, it was worth writing down.
Su's message wasn't about a new chip or a new benchmark. It was about where she thinks the entire industry is headed, and why AMD's positioning right now looks different from where it was even 12 months ago.
Speaking in a Yahoo Finance interview at the conference, Su said AI is at the point where more people are doing genuinely useful, "meaningful work" with it. Not demos. Not pilots. Real production deployments, running 24 hours a day, embedded in business workflows that companies are betting real money on, according to Yahoo Finance.
Her framing is worth unpacking. The early phase of AI was dominated by training, building large models that required enormous, concentrated bursts of GPU power running for months at a time.
Related: Bank of America revamps AMD stock price target for 2026
Inference is different. Inference is every time someone uses an AI product. Every query to a chatbot, every AI agent processing a document, every customer service interaction handled automatically. When AI gets embedded in daily business processes at scale, the required compute volume can grow exponentially.
Su said 2026 marks a historical milestone. It's the first year that global inference compute is projected to surpass training compute.
AMD's internal data back that up. Monthly AI token consumption has reached roughly 35 quadrillion tokens, representing around 160 times growth in just two years, according to CryptoBriefing. That's not a technology story anymore. It's a business story.
One of the most important things Su said is that AI is no longer just a GPU story. Agentic AI systems, which run complex multi-step tasks autonomously, require significant CPU resources for orchestration, data handling, and coordination alongside GPU-based model execution. The two work together, and the demand for both is rising at the same time.
AMD's own numbers tell that story. The company has revised its server CPU market estimate up to $220 billion by 2030, from a prior $120 billion.
Workers just sent AI companies an ultimatum
Palantir CEO has a blunt verdict on OpenAI and Anthropic
Elon Musk pulls no punches with AI rivals as Grok 4.5 debuts
That's not a rounding error. That's AMD saying the CPU market is nearly twice as big as it thought it was, because AI changed what CPUs are needed for. Total AI compute, in AMD's view, is heading toward a $2 trillion opportunity by 2030 at a 40% annual growth rate, with inference driving most of it.
AMD launched its EPYC Venice processor commercially at the event. Venice runs on TSMC's 2nm architecture and is the first chip where that CPU market revision starts showing up in actual product.
Last quarter, AMD's data center segment pulled in $5.8 billion, up 57% year on year. Q2 earnings land Aug. 4. Wall Street is looking for around $11.3 billion in revenue, up 47% from a year ago.
The stock is up roughly 115% this year, even after dropping 8.85% on July 28, when chip stocks got hit broadly on macro concerns and the SK Hynix earnings miss.
Su's inference argument isn't just a vision. AMD has been building the customer relationships to back it up.
In October 2025, AMD announced a partnership with OpenAI involving 6 gigawatts of GPU capacity, with a first phase of 1GW MI450 deploying in the second half of 2026. In February 2026, it signed a similar deal with Meta.
In July 2026, AMD confirmed the Anthropic partnership, involving up to 2 gigawatts of Instinct MI455X GPUs, part of the MI450 Series, integrated into Helios rack-scale solutions alongside EPYC Venice CPUs, Pensando networking, and ROCm software. The first gigawatt deploys in H1 2027, according to CNBC.
The engineering side of the deal is arguably more important than the hardware numbers. Both companies will use Claude to optimize AMD Instinct GPU workloads and accelerate ROCm software development.
That directly targets the gap between AMD's ROCm stack and Nvidia's CUDA ecosystem, which has been AMD's most persistent competitive disadvantage in AI. Getting Anthropic's engineering team to help close that gap is a different kind of win than a purchasing agreement.
AMD is also making an equity investment of up to $5 billion in Anthropic, milestone-contingent on deployment targets. Helios rack systems are priced at $5 million to $5.5 million per rack, reflecting the shift from selling individual GPU components to selling full rack-scale AI systems with significantly higher revenue per unit.
These deals matter beyond the revenue lines. They signal that the three most consequential AI model developers in the market, the companies whose infrastructure decisions shape the entire semiconductor industry, are willing to commit serious capital to AMD's AI stack as an alternative to Nvidia.
AMD dropped 8.85% on July 28. It wasn't anything AMD did. The SK Hynix earnings miss, a hawkish Fed hold, and general chip sector nerves pulled the whole group down.
AMD is still up roughly 115% for the year. Options traders are pricing in a 12.28% move in either direction around Aug. 4. That's a big swing expectation for a stock already up this much.
Goldman Sachs, KeyBanc, UBS, and Mizuho have all raised their AMD price targets in recent weeks. KeyBanc carries a $725 target, UBS a $700 target, and Mizuho recently lifted its target to $625.
Bank of America raised its target to $620 from $560 on July 25, arguing AMD's AI window is opening wider than the market appreciates, as TheStreet reported. Barclays made a similar case in June, raising its target to $665 from $500 on the argument that the AI trade is missing the CPU story entirely, as TheStreet reported. The Street consensus is a Strong Buy with 28 Buy ratings versus eight Holds.
The broader question for investors isn't whether AI demand is real. Su's data on token consumption growth makes that case clearly enough.
The question is whether AMD can continue translating that demand into revenue and margin as the market shifts from training-centric infrastructure to inference-at-scale. The Anthropic, OpenAI, and Meta deals suggest the hyperscalers think it can. Aug. 4 is when the quarterly numbers will either confirm or complicate that thesis.
Data center GPU revenue guidance. Wells Fargo's above-consensus estimate for AMD data center GPU revenue is $40.6 billion in 2027. Any commentary from Su on MI450 demand and Helios system deployments will either strengthen or challenge that projection. The gap between Wells Fargo's estimate and the Street consensus is the key number to watch.
EPYC Venice ramp contribution. The Zen 6 EPYC Venice processor on TSMC's 2nm node launched commercially in July. Q2 is the first quarter where Venice begins contributing meaningfully to server CPU revenue. How quickly that ramp shows up in the data center segment numbers will tell investors whether the $220 billion CPU market thesis is tracking ahead of or behind schedule.
Anthropic deployment timeline update. The first gigawatt of AMD MI450 compute for Anthropic is targeted for the first half of 2027. Any update to that timeline, earlier or later, will move the stock. Earlier means faster revenue recognition. Later raises questions about execution against the milestone-contingent $5 billion equity commitment.
Related: UBS hurries to reset AMD stock target on key AI Day signals
This story was originally published by TheStreet on Jul 31, 2026, where it first appeared in the Investing section. Add TheStreet as a Preferred Source by clicking here.