Nvidia spent two decades convincing the world it builds the best graphics chips on the planet. Now it wants data center buyers to believe it builds the best CPUs too. Nvidia’s press materials first detailed the chip’s specs in March 2026, and on May 31, 2026, at the GTC Taipei keynote, Nvidia formally unveiled Vera, its first fully custom-designed data center processor, branding it “the CPU for Agents.” The chip packs 88 in-house “Olympus” cores exposing 176 threads, pairs with Nvidia’s new Rubin GPU inside the Vera Rubin platform, and goes straight after server territory that Intel and AMD have split between them for decades.
Nvidia CFO Colette Kress told investors the Nvidia Vera CPU opens a $200 billion market the company has never addressed before, with visibility into nearly $20 billion of CPU revenue already booked for this year. Intel still ships around 60% of server CPUs and AMD holds another 24.3%, according to Mercury Research data cited by CNBC, and CNBC’s follow-up reporting in July confirmed Vera is now positioned as a direct challenge to both rivals. Nvidia is starting from just 6.2%. Here is what the specs, the benchmarks, and the money actually show about an 88-core Arm chip built by a GPU company.
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 Is Nvidia’s Vera CPU?
Vera is Nvidia’s first CPU built entirely from its own core design rather than licensed from Arm’s Neoverse line, which is the approach Nvidia used for its earlier Grace CPU. Nvidia made the formal unveiling at the GTC Taipei keynote, and according to Nvidia’s product page, Vera uses 88 custom “Olympus” cores that are instruction-set compatible with Arm v9.2 but architecturally distinct from anything Arm itself sells.
Each chip exposes those 176 threads through a feature Nvidia calls Spatial Multithreading, or SMT-X. Nvidia’s product page confirms, as of June 2026, that the chip supports up to 1.2 TB/s of memory bandwidth and up to 1.5 TB of LPDDR5X capacity per socket, running at a configurable 250W to 450W TDP — figures Tech Insider verified against Nvidia’s own specs in August 2026 and that DigitalApplied independently documented that same June 2026. Nvidia detailed the same 88-core, 176-thread SKU at its Hot Chips 2026 talk, specifying that the 1.2 TB/s bandwidth figure runs through SOCAMM2-format LPDDR5X modules, and QZ’s July 2026 analysis added that Vera’s internal coherency fabric moves data at 3.4 TB/s into a 164 MB unified L3 cache shared across all 88 cores. The Register separately reported in August 2026 that Vera’s 1.2 TB/s of memory bandwidth runs roughly double what AMD’s Turin-generation EPYC chips deliver.
The Nvidia Vera CPU does not work alone. It pairs with Nvidia’s new Rubin GPU over sixth-generation NVLink, sharing a coherent memory address space through NVLink-C2C that The Register, reporting on Vera’s Superchip interconnect in August 2026, measured at 1.8 TB/s. That link matters more than it sounds. It means the CPU and GPU can read each other’s memory directly instead of shuttling data back and forth, which is the bottleneck Nvidia says it built Vera to remove. Nvidia also unveiled a CPU-only Vera Superchip in March 2026 that links two Vera dies over that same 1.8 TB/s NVLink-C2C fabric, combining for 176 cores and 352 threads in a single module, and it sells standalone Vera-only racks that scale up to 256 liquid-cooled chips for buyers who want the CPU without a GPU pairing, according to Tom’s Hardware.
Commercial availability is set for fall 2026, shipping through system builders and cloud partners, though Nvidia’s own delivery blog said in August 2026 that Vera had already entered full production, citing internal benchmarks showing the chip running up to 1.8x faster per core than rival x86 CPUs on agentic AI workloads. Nvidia has not disclosed official pricing for the chip itself.
Inside the Olympus Core and Spatial Multithreading
The Olympus Core: Built From Scratch
The core Nvidia built for Vera has a name of its own: Olympus. It is the company’s first ground-up CPU core design, breaking from the approach Nvidia used with Grace, which relied on Arm’s off-the-shelf Neoverse cores. Nvidia first flagged the 88-core, Arm v9.2-compatible Olympus design in its Rubin-generation roadmap in January 2026, months ahead of the formal GTC Taipei unveiling. Olympus keeps Arm v9.2 compatibility so existing software still runs, but Nvidia designed the fetch, decode, and branch-prediction logic in house, tuned for the branchy, low-arithmetic code that AI agent orchestration produces rather than the dense math GPUs already handle well.
Spatial Multithreading, Not Time-Slicing
Intel’s Hyper-Threading and AMD’s SMT both let two threads share a core by taking turns on the same execution resources. Nvidia’s Spatial Multithreading works differently. It physically partitions a core’s resources so two threads run with real isolation instead of splitting time. Nvidia argues that distinction matters for AI factories running thousands of concurrent agent sandboxes on one socket, where one noisy tenant slowing down another becomes a real operational problem. Whether SMT-X holds up under production workloads at scale is still unproven outside Nvidia’s own labs, since general customers have not received the chip yet.
Vera Rubin NVL72: The Rack-Scale Numbers
Nvidia doesn’t expect most AI factory customers to buy Vera chip by chip. The flagship deployment is the Vera Rubin NVL72 rack, code-named Oberon, which packs 72 Rubin GPUs and 36 Vera CPUs into one liquid-cooled cabinet. Two Rubin GPUs and one Vera CPU sit together on a board Nvidia calls a Superchip, and 36 of those boards make up a full rack.
Each Rubin GPU carries 288 GB of HBM4 memory, and each Vera CPU carries 1.5 TB of LPDDR5X. Multiply that across a full rack and the system holds 20.7 TB of HBM4 and roughly 54 TB of LPDDR5X, connected over sixth-generation NVLink Switch. Nvidia states the rack delivers up to 3,600 petaflops of NVFP4 inference performance, though that figure comes from Nvidia’s own materials and has not been independently verified.
Morgan Stanley Research, in a bill-of-materials estimate reported by WccfTech on May 21, 2026, priced a fully configured NVL72 VR200 rack at roughly $7.8 million. Vera Rubin entered production with first shipments targeted for the third quarter of 2026 and a volume ramp planned for the fourth quarter.
Nvidia vs. Intel vs. AMD: Server CPU Market Share Today
The Nvidia Vera CPU enters a market Intel and AMD have split between them for years. Mercury Research data cited by CNBC put server CPU shipment share at Intel 60%, AMD 24.3%, and Nvidia 6.2% for the fourth quarter of 2025, the most recent full-quarter snapshot before Vera’s launch. The remaining share splits across in-house Arm chips built by Amazon, Microsoft, and Google for their own cloud fleets.
| Vendor | Server CPU Share (Q4 2025) | Core Architecture | Flagship Chip |
|---|---|---|---|
| Intel | 60% | x86 | Xeon 6980P |
| AMD | 24.3% | x86 | EPYC 9575F |
| Nvidia | 6.2% | Arm (custom Olympus core) | Grace / Vera |
| Hyperscaler in-house (AWS, Microsoft, Google) | ~9.5% (remainder) | Arm | Graviton and others |
That 6.2% figure understates what Nvidia is chasing. Kress told investors Nvidia now has visibility into nearly $20 billion of total CPU revenue this year, and an analyst estimate reported by Tom’s Hardware puts Nvidia on track to capture as much as two-thirds of the x86 server CPU market currently split between Intel and AMD, with roughly 4 million Vera CPUs expected to ship in fiscal 2027. Intel and AMD have not lost that share yet, but Nvidia is the only company with a credible seat at both the CPU and GPU side of the table, and it is using that position aggressively.
For context, AMD’s own EPYC Venice line recently posted a record 46.2% of x86 server share, a separate measurement this site covered in June 2026. That underscores how much ground AMD alone has taken from Intel before Nvidia even entered the fight with a full CPU lineup of its own.
What a Vera Rubin Rack Actually Costs
Nvidia has not published Vera pricing, but Morgan Stanley’s bill-of-materials model gives the clearest outside look at where the money goes inside a Vera Rubin rack. Memory is the story. Combined HBM4 and LPDDR5X costs jumped 435% to more than $2 million per rack, up from roughly $373,939 on the prior Grace Blackwell generation.
| Component | Estimated Cost per Rack | Change vs. Grace Blackwell |
|---|---|---|
| Rubin GPUs (72 per rack) | ~$4,000,000 | +57% |
| HBM4 + LPDDR5X memory | ~$2,000,000+ | +435% |
| Vera CPUs (36 per rack) | ~$180,000 (~$5,000/chip) | New product line |
| NVLink, networking, cooling, PCB, other | ~$1,600,000-2,000,000 | PCB cost alone +233% |
| Total estimated rack cost | ~$7,800,000 | N/A |
Memory now accounts for 26% of total rack cost, up from 9% on Grace Blackwell systems, according to the same estimate. Rubin GPUs remain the single biggest line item at close to $4 million per rack, a 57% jump over the prior generation’s roughly $2.5 million. The 36 Vera CPUs in a full rack add up to an estimated $180,000, or about $5,000 per chip. That’s a small slice of total rack cost, but a real one for a product line that did not exist a year ago.
The Performance Claims, and What Independent Testing Found
Nvidia’s own marketing leans hard on multiples of ten. The company claims up to 10x agent throughput, 10x tokens delivered per megawatt, and 10x lower cost per token compared with the Grace Blackwell generation, measured at full pod scale. Nvidia’s March 2026 materials separately claimed Vera runs 50% faster and roughly 2x more power-efficient than traditional rack-scale CPU designs, and its AI Innovators blog updated that pitch in July 2026 to claim 50% higher IPC than Grace. Every one of those figures comes from Nvidia’s internal testing and had not been independently replicated as of early June 2026.
Independent numbers exist, but with a catch. Phoronix ran its own benchmark suite against the 88-core Vera and found it roughly 55% faster than Intel’s best single-socket Xeon 6980P and about 10% to 11% ahead of AMD’s EPYC 9575F, measured in geometric mean across the test suite. Vera also beat Nvidia’s own outgoing Grace CPU by around 1.6x, and SemiAnalysis separately estimated the generational jump from Grace to Vera at roughly 2x performance with chip-to-chip bandwidth doubling to 1.8 TB/s.
The catch: Nvidia chose which workloads Phoronix could run, a mix weighted toward Python execution, code compilation, Java, database operations, and memory-stream tests. General-purpose and legacy enterprise workloads, the kind that still make up a large share of real-world Xeon and EPYC deployments, were left out of the test set. That does not make the benchmark numbers false. It means they describe Vera’s best case rather than its average case, and buyers evaluating a switch from Intel or AMD should treat the 55% figure as a ceiling rather than a guarantee.
Why CPUs Suddenly Matter in the Age of AI Agents
For most of the last decade, CPU choice in AI infrastructure was an afterthought. GPUs did the math that mattered, and the CPU’s job was mostly to keep them fed. Agentic AI changes that math. An agent doesn’t just generate tokens. It plans steps, calls external tools, spins up code sandboxes, queries retrieval systems, and stitches results back together, often dozens of times per user request across hundreds of sessions running at once. Nearly all of that work runs on the CPU, not the GPU.
“CPUs are becoming the bottleneck in terms of growing out this AI and agentic workflow.”
Dion Harris, Head of AI Infrastructure, Nvidia — CNBC, March 13, 2026
That framing is self-serving, since Nvidia sells the fix, but the underlying diagnosis matches what infrastructure teams have reported all year: expensive GPUs sitting idle while a CPU works through a queue of tool calls and sandbox executions. The Nvidia Vera CPU’s entire design, from the Olympus core’s branch prediction to the SMT-X threading model, targets that specific gap rather than trying to out-math a GPU at anything.
Nvidia’s $20 Billion Bet and Wall Street’s Reaction
On Nvidia’s most recent earnings call, Kress laid out the financial case for Vera in blunt terms.
“Vera CPU opens a brand-new $200 billion TAM for Nvidia, a market we have never addressed before.”
Colette Kress, Chief Financial Officer, Nvidia — Tom’s Hardware
She followed with a specific revenue number: “We have visibility to nearly $20 billion in total CPU revenue this year, setting us up to become the world’s leading CPU supplier.” That last claim is aggressive. Intel alone shipped roughly 60% of server CPUs as recently as Q4 2025, and Intel’s data center revenue dwarfs $20 billion on its own. But Kress was not just talking about ambition. “Every major hyperscale and system maker is partnering with us to get it deployed,” she said, tying the revenue number to specific customer commitments rather than a hopeful projection.
Tom’s Hardware reported that Nvidia is already on track to deliver 4 million Vera CPUs in fiscal year 2027, citing an analyst estimate that Nvidia could eventually capture two-thirds of the x86 server market currently split between Intel and AMD. That scenario assumes Vera’s Arm-based design displaces existing x86 sockets rather than simply adding new AI-specific capacity alongside them, which is a meaningfully different, and harder, outcome for Nvidia to achieve.
Who’s Actually Buying the Vera CPU
Nvidia sent pre-production Vera hardware to a small group of evaluators in mid-May 2026, ahead of the broader fall 2026 commercial launch. Anthropic, OpenAI, and Oracle Cloud Infrastructure all received early units for testing around the time of the GTC Taipei launch. OCI has been the most specific about its plans, saying it intends to deploy hundreds of thousands of Vera-based units beginning this year. SpaceXAI has since become one of the first named production deployments, with its own August 2026 rollout notes citing up to 1.8x faster task completion across Vera’s 88 cores running at the chip’s full 1.2 TB/s memory bandwidth.
The interest is not limited to AI labs and cloud providers. The New York Stock Exchange is reportedly evaluating Vera for high-throughput market data processing, a workload that has nothing to do with AI agents but shares Vera’s appetite for fast, branchy, low-latency computation. That is a useful reminder that Nvidia is not only pitching Vera as an AI chip. It is pitching Olympus core performance as a general argument against x86, in any workload where memory bandwidth and thread isolation matter more than raw clock speed.
None of these relationships are binding purchase orders in the way a public earnings report would disclose, and Nvidia has not named production-scale customers beyond these evaluation partners.
From Grace to Vera: Nvidia’s CPU Learning Curve
Vera is not Nvidia’s first attempt at a data center CPU. That was Grace, which launched inside the Grace Hopper Superchip starting in 2023 and later inside Grace Blackwell systems, built on licensed Arm Neoverse cores rather than a custom design, topping out at 72 cores per chip. Grace proved Nvidia could ship a CPU that worked well enough to pair with its own GPUs at scale, but it never seriously threatened Intel or AMD’s shipment volumes. Vera’s monolithic 88-core Olympus die, which ServeTheHome detailed in a July 2026 teardown, adds 16 more cores while dropping licensed Arm IP entirely. The 6.2% server CPU share Mercury Research recorded for Nvidia in Q4 2025 is largely Grace’s legacy.
Vera is a different kind of bet. Instead of licensing Arm’s core design the way Grace did, and the way AWS Graviton and Microsoft’s own Arm chips do, Nvidia built Olympus from scratch. SemiAnalysis estimated the jump from Grace to Vera at roughly 2x performance, with chip-to-chip bandwidth doubling to 1.8 TB/s, and The Next Platform’s January 2026 analysis put the memory-capacity leap at 3.2x, with per-socket LPDDR5X climbing from Grace’s 480 GB ceiling to Vera’s 1.5 TB. That is a bigger architectural leap than most companies attempt in one generation, and it signals Nvidia sees Vera as a platform to build on for years rather than a one-off answer to AMD and Intel’s current roadmaps.
How Intel and AMD Are Positioned to Respond
Neither Intel nor AMD has made a direct public statement responding to Vera by name, based on available reporting. But both companies have moved on parallel tracks that shape how much room Vera actually has to grow.
Intel is mid-turnaround. Its 18A-P manufacturing process entered risk production in mid-2026, and the company has reportedly drawn renewed interest from Apple and Google as foundry customers, alongside a partnership with Taiwan’s UMC to expand capacity. A stronger Intel foundry business does not stop Vera, but it strengthens Intel’s ability to fund competitive Xeon designs instead of losing further ground on process technology.
AMD’s position is more immediately relevant to Nvidia’s pitch. AMD has spent the past two years posting record x86 server share gains with its EPYC line, and it already sells its own GPU-CPU platforms through the MI350 series paired with EPYC processors, a similar basic pitch to what Nvidia is making with Vera Rubin. AMD’s advantage is an existing x86 software base that does not require anyone to recompile for a new Arm-compatible core. Qualcomm is also circling this market from a different angle with its Dragonfly C1000 data center CPU, adding a third Arm-based challenger alongside Nvidia. Nvidia’s real advantage is that Rubin GPUs, not Vera CPUs, are still what most AI buyers actually want, which gives Nvidia leverage Intel and AMD don’t have: it can bundle the CPU nobody asked for with the GPU everybody wants.
What This Means for the Broader AI Hardware Market
Vera’s real significance is not whether it takes 5% or 15% of server CPU shipments by 2027. It is what Nvidia’s entry does to the rest of the market’s assumptions. For most of the AI boom, buyers treated CPU choice as a secondary decision made after picking a GPU vendor. Vera collapses that into a single decision, and every hyperscaler now has to weigh whether a tightly coupled Nvidia CPU-GPU stack is worth trading away the flexibility of mixing Intel or AMD CPUs with Nvidia GPUs.
That pressure compounds with the memory market. Vera Rubin racks already show memory eating 26% of total system cost, up from 9% a generation earlier, and Nvidia’s $500 billion memory supply agreement with SK Hynix only gets more important as Vera adds LPDDR5X demand on top of existing HBM4 demand for Rubin GPUs. Every additional Vera CPU sold is also additional memory demand stacked on a supply chain that is already tight. Data center operators evaluating Vera Rubin racks in 2026 are effectively betting on Nvidia’s ability to secure memory supply as much as betting on the Olympus core’s performance.
Five Predictions for Vera Through 2027
Based on current shipment targets, revenue guidance, and competitive dynamics, here is how the Nvidia Vera CPU story likely plays out over the next 18 months:
- Nvidia lands in the high single digits of total server CPU share by the end of 2027, short of Kress’s most ambitious framing but well above its current 6.2% base, driven mostly by bundled Vera Rubin rack sales rather than standalone CPU purchases.
- Intel and AMD lean harder into tighter CPU-accelerator packaging rather than attempt a custom Arm core of their own, since neither company has signaled a ground-up core redesign on Vera’s timeline.
- Independent benchmark coverage expands once Vera ships to general customers in fall 2026, and some of Nvidia’s 10x pod-scale claims narrow once tested outside Nvidia-selected workloads.
- HBM4 and LPDDR5X supply, not core design, becomes the binding constraint on how many Vera Rubin racks actually ship in 2027, keeping rack prices elevated industry-wide.
- Vera adoption spreads outside AI entirely, following NYSE’s interest in market-data processing, as a signal of whether Olympus cores can compete on general performance rather than only on AI-agent workloads.
Frequently Asked Questions About Nvidia’s Vera CPU
What is Nvidia’s Vera CPU?
Vera is Nvidia’s first fully custom-designed data center CPU, unveiled at GTC Taipei on May 31, 2026. It uses 88 in-house “Olympus” cores and is built to pair with Nvidia’s Rubin GPU inside the Vera Rubin platform.
How many cores and threads does Vera have?
Vera has 88 physical cores and exposes 176 threads through Nvidia’s Spatial Multithreading (SMT-X) technology, which partitions core resources rather than time-slicing them.
Is the Vera CPU based on Arm architecture?
Yes. Vera is Arm v9.2 instruction-set compatible, but unlike Nvidia’s earlier Grace CPU, it uses a custom Nvidia-designed core called Olympus rather than licensed Arm Neoverse cores.
How does Vera perform against Intel Xeon and AMD EPYC?
Independent Phoronix testing found Vera roughly 55% faster than Intel’s Xeon 6980P and about 10% to 11% ahead of AMD’s EPYC 9575F in geometric mean, though Nvidia selected the benchmark workloads, so general-purpose performance remains untested.
When can I buy a Vera CPU, and how much does it cost?
Nvidia has set commercial availability for fall 2026 through system builders and cloud partners. The company has not disclosed official pricing, though Morgan Stanley Research estimates put Vera CPUs at roughly $5,000 per chip inside a full Vera Rubin rack.
What is the Vera Rubin NVL72 platform?
It is Nvidia’s rack-scale system pairing 36 Vera CPUs with 72 Rubin GPUs over sixth-generation NVLink, with a coherent shared memory space between CPU and GPU. Morgan Stanley estimates a fully configured rack costs about $7.8 million.
Which companies are testing Vera CPUs?
Anthropic, OpenAI, and Oracle Cloud Infrastructure received pre-production Vera hardware in mid-May 2026. OCI has said it plans to deploy hundreds of thousands of units starting this year, and the New York Stock Exchange is separately evaluating Vera for market-data workloads.
Can I buy Vera CPUs without Rubin GPUs?
Yes. Nvidia offers standalone Vera racks that scale up to 256 liquid-cooled CPUs for customers who want Olympus-core performance without pairing it to Rubin GPUs.
The Bottom Line
Nvidia did not build the Vera CPU to beat Intel and AMD at their own game. It built a chip shaped entirely around the workload Nvidia already dominates: AI infrastructure. Whether that is enough to move server CPU market share in any meaningful way by 2027 depends on memory supply, independent benchmark scrutiny, and whether hyperscalers are willing to trade x86 flexibility for tighter Nvidia integration. The 6.2% starting line is real. So is the $20 billion revenue number Nvidia’s own CFO put on the table.
Related Coverage
- AMD EPYC Venice Hits Record 46.2% Server Share [2026]
- Qualcomm’s Dragonfly C1000 Lands Meta, Eyes $15B [2026]
- AWS Graviton vs Intel/AMD: 45% Cheaper, 25% Faster [2026]
- AMD MI350P vs Nvidia H200: 40% Faster, $30K Price [2026]
- Nvidia, SK Hynix Ink $500B AI Memory Deal [2026]
- Intel Hikes Core Ultra Plus CPU Prices Up to 17% [2026]
