AMD used its Advancing AI 2026 conference at San Francisco’s Moscone Center to move from roadmap slides to shipping silicon. On July 23, CEO Lisa Su walked a packed keynote hall through the first volume-production details of the MI400 accelerator family, the Helios rack-scale system built around it, and EPYC Venice, the industry’s first x86 server processor built on TSMC’s 2nm node. The announcements landed alongside confirmation that OpenAI and Meta have combined for 12 gigawatts of AMD accelerator capacity, with Microsoft Azure and Oracle named as early Helios customers.
AMD shares traded near $553 during the two-day event, more than double where they started 2026, as Wall Street analysts raised price targets into the $600s and $700s. The event marks AMD’s clearest attempt yet to convert Nvidia’s AI infrastructure customers into multi-vendor buyers, backed by a rack that AMD says packs up to 432GB of HBM4 per GPU across its 72 MI455X accelerators. Here’s what actually shipped, what it costs, and what it means for the AI hardware market heading into 2027.
Don't miss new tech stories on Google
Add Tech Insider once in the Google app and our stories appear in your news suggestions.
What Happened at AMD Advancing AI 2026
AMD’s flagship AI event ran July 22-23, 2026, at the Moscone Center in San Francisco, with Su’s keynote starting at 9:30 a.m. Pacific on the second day. AMD called it its flagship global AI event, and the program backed that up with more than 100 sessions spanning keynotes, an exhibitor floor, workshops, and technical deep-dives aimed at developers, enterprise buyers, and cloud partners. OpenAI CEO Sam Altman joined Su on stage, a symbolic moment given OpenAI’s position as one of AMD’s largest disclosed AI customers.
The core news out of Advancing AI 2026 broke into three buckets: silicon (the MI400 series and EPYC Venice), systems (Helios), and software (ROCm 7). AMD had previewed pieces of this roadmap at CES 2026 in January, when it first detailed the Helios concept and MI400 lineup on paper. Advancing AI 2026 is where those previews turned into confirmed volume-production timelines, rack pricing, and named customers, which is why the stock and analyst reaction differed so sharply from the muted response to the January preview.
Inside the MI400 Family: MI430X, MI440X, and MI455X
AMD’s MI400 series ships as three distinct parts built on the same CDNA 5 architecture and HBM4 memory subsystem, each aimed at a different buyer. The MI430X targets sovereign AI programs, high-performance computing centers, and hybrid CPU-GPU workloads where government and research customers need control over where their models run. The MI440X packages eight GPUs with a single EPYC Venice CPU into a rack-mounted server built for enterprises that want on-premises training, fine-tuning, and inference without renting cloud capacity.
The MI455X is the flagship, and it’s the part that actually goes into Helios. Each MI455X carries 432GB of HBM4 memory and 19.6 TB/s of memory bandwidth per GPU, roughly 50% more capacity than the 288GB Nvidia packs into each Vera Rubin NVL144 GPU. All three chips share AMD’s Infinity Fabric interconnect and add support for UALink, the open scale-up standard AMD has pushed as a counterweight to Nvidia’s proprietary NVLink. Across the family, AMD claims up to 10x more performance for frontier-model workloads compared with its prior generation, with a full Helios rack rated at 2.9 exaflops of FP4 inference performance.
| Chip / System | Generation | Memory | Peak Rack Performance | Status |
|---|---|---|---|---|
| MI355X | Current (CDNA 4) | 288GB HBM3E | 40% more tokens/$ vs. prior chip | Shipping now |
| MI430X | MI400 series | HBM4 | Part of 2.9 exaflops FP4/rack | Launching H2 2026 |
| MI440X | MI400 series | HBM4, 8-GPU server | Part of 2.9 exaflops FP4/rack | Launching H2 2026 |
| MI455X | MI400 series (Helios flagship) | 432GB HBM4/GPU | 19.6 TB/s bandwidth/GPU | Launching H2 2026 |
| Helios (72x MI455X rack) | Rack-scale system | 31TB HBM4 total | 2.9 exaflops FP4 / 1.4 exaflops FP8 | In production; ships late Q3 2026, ~$5.25M |
| MI500 series | Next-gen (2027) | Not yet disclosed | Up to 1,000x MI300X (claimed) | Announced, 2027 target |
Helios: AMD’s First Rack-Scale AI System
Helios is the bigger story. It’s AMD’s first attempt at shipping a complete rack-scale system rather than selling individual accelerators for someone else to integrate, and Su called it “the world’s best AI rack” on stage. A single Helios rack packs 72 MI455X accelerators alongside EPYC Venice CPUs and AMD’s Pensando networking silicon, adding up to 31TB of pooled HBM4 memory and 1.4PB/s of aggregate memory bandwidth. AMD rates the rack at 2.9 exaflops of FP4 inference throughput and 1.4 exaflops of FP8 training throughput, with 260TB/s of scale-up interconnect bandwidth and 43TB/s of Ethernet scale-out bandwidth tying racks together into larger clusters.
“Now, at the heart of Helios, though, is the MI400 series. This is truly the most advanced accelerator we’ve ever built. It’s really engine for the next generation of AI, and it’s designed to run trillion plus parameter models.”
Lisa Su, CEO of AMD — AMD Advancing AI 2026 keynote
That last line is a nod to the size of frontier models now in training at OpenAI, Meta, and other labs. AMD confirmed a double-wide Helios variant arriving in the third quarter of 2026 that pushes a single rack to 3 AI exaflops, which Su described as a blueprint for “yotta-scale compute.” As of August 2026, AMD says the standard Helios rack is in full production, with first customer shipments still scheduled for late Q3 2026.
Helios Production Status: What’s Confirmed for Q3 2026
Advancing AI 2026 answered the question every hyperscale buyer actually cares about: is this real hardware or another roadmap slide? Heading into AMD’s August 4 earnings report, here’s where Helios stands.
- Production status: AMD says Helios has moved out of engineering samples and into full production.
- First shipments: Customer racks are scheduled to ship starting late Q3 2026, inside the second-half window AMD had already guided to.
- Booked demand: AMD points to 12 GW of AI infrastructure demand tied to the platform, anchored by four named customers: Microsoft, Meta, OpenAI, and Oracle.
- Pricing: Unchanged since the event’s initial disclosure, at $5.0 million to $5.5 million per rack, averaging about $5.25 million.
None of this is independently audited. The 12-gigawatt figure is AMD’s own characterization of customer commitments, gigawatts don’t translate cleanly into a specific GPU or rack count, and “full production” is a company claim rather than a shipment AMD has publicly documented leaving a factory. Still, it’s the most concrete production and shipping commitment AMD has attached a date to since the January 2026 preview.
August 2026: The Countdown to Q3 Shipments
AMD enters its August 4 second-quarter earnings report with Helios still roughly two months from its guided late-Q3 2026 shipment window. Wall Street consensus calls for $11.2 billion in quarterly revenue, up 46% year over year, and non-GAAP earnings of $1.67 per share. The print is the first real opportunity for AMD to show whether the 12 GW of booked demand and $5.25 million average rack pricing disclosed at Advancing AI 2026 are converting into backlog and forward guidance, rather than staying keynote commitments.
- Helios commentary on the call: Analysts are expected to press AMD for any update on production yield, customer readiness, or slippage risk inside the late-Q3 2026 window.
- Additional customer names: Microsoft, Meta, OpenAI, and Oracle remain the only four hyperscalers AMD has named against the 12 GW figure; whether a fifth commitment surfaces is one of the open questions heading into the report.
- No pricing or demand changes yet: As of early August 2026, AMD has not revised the $5.0 million to $5.5 million rack pricing or the 12 GW demand figure disclosed at the event.
EPYC Venice Becomes the First 2nm x86 Server Chip
The CPU side of Advancing AI 2026 belonged to EPYC Venice, AMD’s Zen 6 server processor and the first x86 server chip in the industry to enter volume production on TSMC’s 2nm process node. AMD says Venice scales up to 256 cores and delivers roughly 1.7x the performance of the prior-generation EPYC Turin lineup, a jump AMD attributes to the combination of the new process node and Zen 6’s redesigned core architecture.
Venice is the CPU half of every Helios rack, pairing with MI455X GPUs to handle orchestration, data movement, and the CPU-bound portions of AI training pipelines that GPUs alone can’t handle efficiently. AMD said Venice begins shipping in the second half of 2026, the same window as Helios and the MI400 GPU family, which lets AMD sell the platform as a single coordinated launch rather than staggering CPU and GPU availability the way some prior generations did.
“We’re already deep into development of our 2027 rack that will push the efficiency and scalability with our next generation Verano CPUs and MI500 GPUs.”
Lisa Su, CEO of AMD — AMD Advancing AI 2026 keynote
That line gave the audience its clearest public signal yet of AMD’s chip naming and timeline beyond 2026, pairing the already-teased MI500 GPU with a new CPU codename, Verano, for the generation after Venice.
ROCm 7 and AMD’s Software Bet
Hardware announcements dominated the headlines, but AMD spent a meaningful chunk of the keynote on ROCm 7, its open-source GPU software stack and the piece of the platform most directly compared with Nvidia’s CUDA. AMD says ROCm 7 delivers 3.5x the performance of ROCm 6 and has deepened integration with open inference frameworks including vLLM, SGLang, and llm-d, three projects widely used across the open-source LLM-serving ecosystem. AMD also introduced ROCm Enterprise AI and an AMD Developer Cloud, aimed at giving smaller developers and enterprises a way to test workloads on Instinct hardware without buying a rack outright.
The software push matters because AMD’s hardware specs have rarely been the bottleneck in its competition with Nvidia. The gap has historically been software maturity: how well frameworks like PyTorch run out of the box, how many pre-tuned kernels exist for common model architectures, and how much engineering time customers have to spend porting CUDA-optimized code. A developer checking whether their environment sees MI400-class hardware after installing ROCm 7 drivers runs the same basic commands AMD has supported for several generations of Instinct hardware:
$ rocm-smi --showproductname
$ rocminfo | grep "Marketing Name"
$ rocm-smi --showmeminfo vramWhether ROCm 7’s claimed 3.5x gain closes the practical software gap with CUDA will show up in independent benchmarks over the next few quarters, not in keynote slides.
The OpenAI Deal: 6 Gigawatts and a 160-Million-Share Warrant
The customer story anchoring Advancing AI 2026 traces back to October 2025, when AMD and OpenAI announced a multi-year, multi-generation partnership to deploy 6 gigawatts of AMD GPU capacity, starting with a 1-gigawatt tranche of MI450-class accelerators in the second half of 2026. The deal’s most unusual feature is a warrant giving OpenAI the right to buy up to 160 million AMD shares at $0.01 each, vesting in tranches tied to deployment milestones running through October 2030.
If OpenAI exercises the warrant in full, it would hold roughly 10% of AMD’s outstanding shares, a stake large enough that AMD’s stock jumped more than 20% the day the deal was first announced, according to CNBC. The structure ties OpenAI’s upside directly to AMD’s execution. OpenAI only gets cheap shares if AMD actually ships the gigawatts of compute it promised, and if AMD’s stock price clears certain targets along the way. Advancing AI 2026 served as the checkpoint where AMD reaffirmed the timeline publicly, with the first OpenAI MI450-class deployments now confirmed for the second half of 2026, the same window as the broader Helios and MI400 launch.
Meta, Microsoft, and Oracle: The 12-Gigawatt Question
OpenAI isn’t AMD’s only hyperscale commitment. Meta has separately confirmed its own 6-gigawatt AMD GPU deployment across multiple chip generations, also starting with roughly 1 gigawatt of MI450-class hardware in the second half of 2026. Combined, Meta and OpenAI now account for 12 gigawatts of committed AMD accelerator demand, though the two deals are structured differently, and only OpenAI’s includes a disclosed share warrant.
Translating gigawatts into GPU counts is imprecise, since power draw per chip varies with generation and cooling design, but industry estimates put one gigawatt at roughly 25,000 to 50,000 high-end GPUs. That would put the Meta and OpenAI commitments in the range of several hundred thousand accelerators once fully deployed. Microsoft Azure confirmed a separate role as an anchor customer for Helios rack-scale systems specifically, planning to deploy the racks for frontier-model inference workloads. Oracle was also named among early Helios adopters.
None of these four customers, Meta, OpenAI, Microsoft, or Oracle, has dropped Nvidia as a supplier. All are treating AMD as a second source for AI compute at a moment when Nvidia GPU allocation remains the primary bottleneck constraining how fast every major AI lab can train and serve models.
What a $5.25 Million Rack Actually Buys
Pricing has been the missing piece of AMD’s AI rack pitch since the January preview, and Advancing AI 2026 filled in the number. A Helios rack costs between $5 million and $5.5 million, averaging around $5.25 million, according to analysts covering the event. That buys 72 MI455X GPUs, the accompanying EPYC Venice CPUs, Pensando networking gear, and the rack-level engineering AMD has done to keep 72 accelerators fed with power and cooling in a single chassis.
For context, a customer buying at that price is paying roughly $73,000 per GPU once the CPUs, networking, and integration are factored in, though AMD hasn’t broken out a separate per-GPU list price. The pricing puts Helios in the same rough neighborhood as Nvidia’s rack-scale GB300 NVL72 and Vera Rubin NVL144 systems, which similarly bundle GPUs, CPUs, and networking into a single SKU rather than selling components separately. That bundling is deliberate on both sides. Hyperscalers buying racks by the dozen care more about total cost per exaflop and per watt than the sticker price of any individual chip, and neither AMD nor Nvidia wants customers price-shopping GPUs against CPUs and networking piecemeal.
AMD Helios vs. Nvidia Vera Rubin NVL144
AMD spent much of Advancing AI 2026 positioning Helios directly against Nvidia’s Vera Rubin NVL144, the rack-scale system Nvidia detailed at its own GTC event earlier in 2026 and plans to ship later this year. The two platforms take different architectural bets: Nvidia packs more total GPUs into its rack, while AMD counters with more memory per GPU and a wider aggregate bandwidth figure.
| Spec | AMD Helios | Nvidia Vera Rubin NVL144 |
|---|---|---|
| GPUs per rack | 72x MI455X | 144 Rubin GPUs (72 packages) |
| CPU | EPYC Venice, up to 256 cores | 36x Vera CPUs, 88-core Arm v9.2 |
| Memory per GPU | 432GB HBM4 | 288GB HBM4 |
| Bandwidth per GPU | 19.6 TB/s | 13 TB/s |
| Aggregate rack bandwidth | 1.4 PB/s | 260 TB/s (NVLink/CX9) |
| FP4 inference (per rack) | 2.9 exaflops | 3.6 exaflops |
| FP8 training (per rack) | 1.4 exaflops | 1.2 exaflops |
| Availability | In production, late Q3 2026 | 2026 |
Nvidia’s rack still wins on raw FP4 inference throughput, but AMD leads on the FP8 training figure and on memory per chip, the spec that matters most for fitting today’s largest models onto fewer GPUs. Neither company has published independent, third-party benchmark results for either rack yet, since both platforms are still ramping toward broad availability in 2026. Every figure in the table above comes from AMD’s and Nvidia’s own disclosures rather than an apples-to-apples lab test.
Wall Street’s Verdict: AMD Stock Rallies to $553
AMD shares closed at $544.43 on July 21, up 8.11% in the stock’s strongest single session since late June, as investors positioned ahead of the conference. By the time Su took the stage on July 23, AMD was trading around $553, more than double its level at the start of 2026 and sitting about 14% below its June 30 high of $584.73. The stock’s run predates Advancing AI 2026 itself. Shares initially jumped more than 20% back in October 2025 when the original OpenAI warrant deal was announced, and have climbed further through 2026 as AMD’s data center revenue accelerated.
Analysts largely used the event to raise price targets rather than change ratings.
| Firm | Price Target | Stance |
|---|---|---|
| KeyBanc | $725 | Bullish |
| UBS | $700 | Bullish |
| Rosenblatt | $655 | Bullish |
| Goldman Sachs | $640 | Bullish |
| Stifel | $635 | Bullish |
| Morgan Stanley | No raised target disclosed | Equal Weight (outlier) |
Morgan Stanley remains the most notable holdout, a contrarian position given that roughly 82.4% of analysts covering AMD rate the stock a buy. AMD’s next test arrives quickly: the company reports second-quarter earnings on August 4, with Wall Street consensus calling for $11.2 billion in revenue, up 46% year over year, and non-GAAP earnings of $1.67 per share.
From MI300X to MI500: How AMD Got Here
AMD’s path to Advancing AI 2026 runs through several generations of Instinct GPUs that steadily closed the gap with Nvidia’s data center lineup. The MI300X, AMD’s first GPU to win meaningful hyperscale adoption, established AMD as a credible second source for AI training and inference starting in 2024.
“Look, the MI350 series delivers just a massive 4x generational leap in AI compute to accelerate both training and inference.”
Lisa Su, CEO of AMD — AMD Advancing AI 2026 keynote
The current MI350 series pushed memory capacity to 288GB per GPU, enough to run models with up to 520 billion parameters on a single accelerator. The MI355X variant within that family claims up to 10x lower inference token cost than Blackwell, Nvidia’s preferred way of framing its price-to-performance case. MI400 and Helios represent the next step up, and AMD has already signaled what comes after: MI500-series GPUs paired with new Verano CPUs, targeted for 2027, that Su said would deliver up to a 1,000x increase in AI performance compared with the original MI300X. That figure spans multiple architecture generations rather than a single-chip comparison, but it captures how fast AMD believes the AI accelerator market is still moving.
What This Means for the AI Chip Supply Chain
The immediate market impact of Advancing AI 2026 isn’t about AMD replacing Nvidia. It’s about AI labs and hyperscalers gaining real leverage in GPU price and supply negotiations for the first time since the current AI buildout began. Nvidia has spent the past three years as the default, often sole, supplier of the accelerators powering frontier model training, giving it pricing power that shows up in Nvidia’s gross margins and in reports of GPU allocation lists that determine which startups can access enough compute to train competitive models.
Every gigawatt AMD ships to OpenAI or Meta is compute that doesn’t have to come from Nvidia, and even a modest double-digit percentage of hyperscale AI workloads shifting to AMD changes the negotiating dynamic for the workloads that stay on Nvidia hardware. AMD’s bet on open standards, UALink for scale-up interconnect and open Ethernet-based scale-out networking, also matters beyond AMD’s own chip sales, since it gives cloud providers and networking vendors a path to build AI infrastructure that isn’t locked into Nvidia’s proprietary NVLink ecosystem.
The AI accelerator market broadly is projected to exceed $500 billion by 2028, with inference workloads growing more than 80% annually as agentic AI systems multiply the number of model calls each application makes. AMD’s entire Advancing AI 2026 pitch was built around capturing a larger share of that inference growth rather than the training market Nvidia still dominates.
5 Predictions for AMD Through 2027
Advancing AI 2026 set the roadmap. Whether AMD hits every mark is a separate question. Here’s what to watch over the next 18 months:
- Helios ramps gradually, not immediately. AMD said Helios ships in the second half of 2026, but rack-scale systems this complex typically see slow initial ramps. Expect Microsoft Azure and Oracle deployments to scale through 2027 rather than hit full volume by year-end.
- AMD’s data center revenue keeps climbing. Q1 2026 data center revenue already reached $5.8 billion, up 57% year over year. Helios and MI400 shipments starting in the second half of 2026 should push that growth rate higher through 2027.
- At least one more frontier lab joins OpenAI and Meta as an AMD customer. Analysts including Jefferies flagged the possibility of an Anthropic partnership ahead of Advancing AI 2026. Even without a deal at this specific event, the pressure on AI labs to secure non-Nvidia compute capacity keeps building.
- MI500 and Verano CPUs arrive in 2027 under intense scrutiny of the 1,000x claim. AMD’s performance claim versus MI300X will get tested against whatever Nvidia’s next-generation Rubin Ultra platform delivers on a similar timeline.
- UALink adoption becomes the real swing factor. Whether AMD’s open interconnect standard attracts networking and cloud partners beyond AMD’s own ecosystem will determine if Helios remains a single-vendor alternative to Nvidia or becomes the foundation of a broader open AI infrastructure standard.
Frequently Asked Questions
What is AMD Advancing AI 2026?
AMD Advancing AI 2026 is AMD’s flagship annual AI event, held July 22-23, 2026, at San Francisco’s Moscone Center, where the company detailed its MI400 GPU family, Helios rack-scale system, EPYC Venice CPUs, and ROCm 7 software stack.
What is AMD Helios?
Helios is AMD’s first complete rack-scale AI system, combining 72 MI455X GPUs, EPYC Venice CPUs, and Pensando networking into a single rack rated at 2.9 exaflops of FP4 inference performance, priced between $5 million and $5.5 million. As of August 2026, AMD says Helios is in full production, with first shipments scheduled for late Q3 2026.
What’s the difference between the MI430X, MI440X, and MI455X?
All three share AMD’s CDNA 5 architecture and HBM4 memory, but target different buyers. The MI430X is built for sovereign AI and HPC customers, the MI440X packages eight GPUs into an on-premises enterprise server, and the MI455X is the flagship chip that powers Helios racks at hyperscale.
How much does an AMD Helios rack cost?
Analysts covering the event put Helios pricing between $5 million and $5.5 million per rack, averaging around $5.25 million, though AMD hasn’t published an official list price.
What is EPYC Venice?
EPYC Venice is AMD’s Zen 6 server CPU, the first x86 server processor built on TSMC’s 2nm process node, offering up to 256 cores and roughly 1.7x the performance of the prior EPYC Turin generation.
How does Helios compare to Nvidia’s Vera Rubin NVL144?
Nvidia’s rack packs more total GPUs, 144 versus 72, and edges out Helios on raw FP4 inference throughput at the rack level, 3.6 exaflops versus 2.9 exaflops. AMD’s MI455X counters with more memory per GPU, 432GB of HBM4 versus 288GB on Nvidia’s Rubin GPUs, and higher per-GPU bandwidth at 19.6 TB/s versus 13 TB/s.
When will Helios and MI400 chips ship?
AMD says Helios has entered full production, with first customer shipments scheduled for late Q3 2026. The broader MI400 chip family ships across the second half of 2026, with a higher-performance double-wide Helios variant also arriving in the third quarter.
What does the OpenAI-AMD warrant deal mean for AMD stock?
OpenAI holds a warrant to buy up to 160 million AMD shares at $0.01 each, vesting through October 2030 as OpenAI deploys up to 6 gigawatts of AMD GPU capacity. The structure ties roughly 10% of AMD’s potential future share count to the success of the partnership.
What’s next for AMD Helios after Advancing AI 2026?
The next milestone is AMD’s second-quarter earnings report on August 4, 2026, where analysts will look for confirmation that the 12 GW of booked Helios demand and $5.25 million average rack pricing are converting into backlog ahead of first customer shipments in late Q3 2026.
Related Coverage
- AMD MI400 Series: $7.2B AI GPU Challenging Nvidia [2026]
- AMD Earnings: $5.8B Data Center, Stock Jumps 14% [2026]
- NVIDIA Rubin GPU: 336B Transistors, T Orders [2026]
- Nvidia Skips New Gaming GPUs, Breaks 30-Year Streak [2026]
- RTX 5090 vs RTX 5070 Ti for AI: 2x VRAM, 2.7x Price [2026]
- AMD 9800X3D vs 9950X3D vs Intel 285K: 25-35% Faster [2026]
- How to Run an LLM Locally: 13 Steps, 90 Min [2026]
