AMD spent four years walking away from the PCIe add-in card as a format for its data center GPUs, betting instead on the OAM modules that power hyperscale AI racks. On May 7, 2026, it walked back. The AMD Instinct MI350P is a 600-watt, dual-slot PCIe card built to slide into the same air-cooled servers running today, and AMD says it beats Nvidia’s H200 NVL by as much as 43% on key AI math. The card carries 144GB of HBM3E memory, a full CDNA 4 die, and a price tag that, by most reporting, lands in the same $30,000-to-$40,000 band as the Nvidia part it’s chasing.
The timing is not incidental. AMD’s data center segment just posted $5.8 billion in quarterly revenue, up 57% year over year, while Nvidia’s data center business alone pulled in $75.2 billion in roughly the same stretch. The MI350P won’t close that gap. But it does something narrower and arguably more useful for AMD right now: it gives enterprise buyers a drop-in alternative to Nvidia’s PCIe silicon without asking them to rebuild a single rack.
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 the AMD Instinct MI350P?
The MI350P is AMD’s first PCIe-form-factor Instinct accelerator since the MI210 launched in 2022 — a four-year gap during which AMD’s data center GPU roadmap ran almost entirely through OAM (OCP Accelerator Module) parts built for liquid-cooled, rack-scale hyperscaler deployments: the MI300X, the MI325X, and most recently the MI350X and MI355X. Those chips compete for hyperscaler wallet share against Nvidia’s H100, H200, and Blackwell-generation GPUs. The MI350P targets a different buyer entirely.
AMD announced the card in a blog post titled “AMD Instinct MI350P PCIe GPUs: Run Enterprise AI on Your Existing Infrastructure,” and that title is the whole pitch. Instead of asking an enterprise IT department to qualify a new liquid-cooling loop or a new rack design, the MI350P is a standard full-height, full-length, 10.5-inch, dual-slot PCIe CEM card that fits into off-the-shelf air-cooled servers. It’s essentially half of an MI350X shrunk to fit PCIe power and thermal limits, built on the same CDNA 4 architecture and the same TSMC 3nm/6nm FinFET process as its OAM siblings.
That positioning matters because most enterprises outside the hyperscaler tier don’t have — and don’t want to build — the liquid-cooled infrastructure that OAM accelerators require. AMD is explicitly chasing the on-premises AI inference market: banks, healthcare systems, manufacturers, and mid-size cloud providers that want to run generative and agentic AI workloads on hardware they already know how to rack, power, and cool.
Full Specs: Inside the MI350P’s CDNA 4 Architecture
AMD’s official numbers, drawn from the product brochure and the company’s own MI350P product page, put the card firmly in the same performance class as Nvidia’s fastest PCIe AI accelerators. The headline spec is 144GB of HBM3E memory spread across four stacks, feeding 128 compute units, 512 matrix cores, and 8,192 stream processors running at a peak engine clock of 2.2GHz.
| Specification | AMD Instinct MI350P |
|---|---|
| Architecture | CDNA 4, TSMC 3nm/6nm FinFET |
| Form factor | FHFL 2-slot PCIe CEM card, 10.5″, fanless/passive |
| Host interface | PCIe Gen 5 x16 |
| Memory capacity | 144GB HBM3E |
| Memory bandwidth | 4 TB/s peak theoretical |
| Compute units / matrix cores / stream processors | 128 / 512 / 8,192 |
| Peak engine clock | 2.2 GHz |
| Last-level cache | 128MB |
| FP16/BF16 peak compute | 1,150 TFLOPS |
| FP8 peak compute | 2,300 TFLOPS |
| FP4/MXFP4 peak compute | 4,600 TFLOPS |
| Board power | 600W, configurable to 450W |
| Announcement date | May 7, 2026 |
The precision-scaling story is the same one AMD has told since the MI300 generation: at full FP16/BF16 precision the card delivers 1,150 TFLOPS, but drop to FP8 and that doubles to 2,300 TFLOPS, and drop further to MXFP4 or MXFP6 — the low-precision formats increasingly used for large language model inference — and peak throughput reaches 4,600 TFLOPS. That scaling is central to AMD’s pitch, since most production inference workloads today run at FP8 or lower, not full precision.
MI350P vs Nvidia H200 NVL: The Benchmark Numbers
AMD built the MI350P to be cross-shopped against one specific part: Nvidia’s H200 NVL, the PCIe-based version of Nvidia’s Hopper-generation flagship. On paper, the two cards trade blows rather than one clearly beating the other.
| Spec | AMD MI350P | Nvidia H200 NVL | Nvidia RTX Pro 6000 Server |
|---|---|---|---|
| Memory capacity | 144GB HBM3E | 141GB HBM3e | 96GB GDDR7 |
| Memory bandwidth | 4.0 TB/s | 4.8 TB/s | 1.6 TB/s |
| Board power | 600W (450W configurable) | 600W configurable | Lower TDP class |
| Form factor | PCIe Gen 5 x16, 2-slot | PCIe / NVL configuration | PCIe |
| AMD-claimed compute edge vs MI350P | Baseline | ~20% behind on FP64, ~43% behind on FP16, ~39% behind on FP8 (AMD figures) | Not directly compared by AMD |
AMD’s own figures claim the MI350P delivers roughly 20% better FP64 performance, 43% better FP16 performance, and 39% better FP8 performance than the H200 NVL — numbers Tom’s Hardware summarized as “roughly 40% faster in FP16 and FP8 theoretical compute.” Independent reporting lands close to that same range: The Register put the FP8 advantage at roughly 38%, writing that “compared to the 141 GB H200, the MI350P promises about 38 percent higher peak performance at FP8, while eking out a narrow VRAM capacity advantage.”
Nvidia isn’t without an answer on paper, though. The H200 NVL still wins on raw memory bandwidth — 4.8 TB/s against the MI350P’s 4 TB/s, a roughly 20% edge that matters for memory-bound inference workloads with long context windows. And all of these are AMD’s own theoretical peak figures rather than independently verified, real-world inference benchmarks, a distinction worth keeping in mind before treating “43% faster” as a guarantee across every workload.
Pricing: What the MI350P Actually Costs
AMD did not publish an official list price at launch — a gap confirmed independently by both Phoronix and ServeTheHome, whose launch-day coverage both note that no pricing was included in the pre-launch briefing materials. That’s not unusual for enterprise accelerators, which are typically sold and priced through OEM server partners rather than at a fixed MSRP.
Market reporting since launch has pegged real-world MI350P pricing in the $30,000-to-$40,000 range per card — squarely in the same band as street pricing for Nvidia’s H200 NVL. That positioning is deliberate: AMD isn’t trying to win this fight on price. It’s trying to win on the argument that a buyer gets more memory, a friendlier power envelope, and a stated compute edge for a comparable dollar figure, while also avoiding Nvidia’s persistent allocation constraints on H200-class hardware.
Where MI350P Pricing and Availability Stand in August 2026
Three months after the May 7, 2026 launch, AMD still hasn’t published official list pricing for the MI350P in its primary marketing materials — the same gap Phoronix and ServeTheHome flagged at launch persists into August 2026. Market coverage from Tom’s Hardware and The Register continues to place real-world pricing in the $30,000-to-$40,000 range, and both outlets frame that figure as roughly comparable to H200-class pricing rather than a clearly cheaper alternative — a more modest cost advantage than the MI350P’s compute benchmarks alone would suggest.
Dell’s PowerEdge XE7745 and R7725 systems, the first MI350P-qualified servers, were slated for availability starting in July 2026. That window has now passed, but AMD and its OEM partners have not published unit shipment figures or an official price list to confirm volume availability — buyers are still pricing MI350P deals through Dell, HPE, Supermicro, and Gigabyte on a quote-by-quote basis rather than against a published MSRP.
Why AMD Went Back to a PCIe Card
The strategic logic behind reviving the PCIe form factor comes down to addressable market. Hyperscalers buying OAM-based MI300X, MI325X, and MI350X/MI355X accelerators represent a small number of very large customers who can justify liquid-cooling retrofits and custom rack designs. Everyone else — the vast tier of enterprises, regional cloud providers, and system integrators — mostly can’t, or won’t, without a multi-year infrastructure project.
A standard PCIe card removes that barrier entirely. Any data center that already runs air-cooled GPU servers for existing workloads can add MI350P cards without touching facilities infrastructure. AMD is explicitly marketing the card’s ROCm software stack as “open, low- and no-cost” compared to Nvidia’s CUDA ecosystem, positioning the total cost of ownership pitch around software licensing as much as silicon.
It’s also a hedge against Nvidia’s chronic supply constraints. Enterprises that have spent the past two years on waitlists for H100 and H200 allocation now have a second real sourcing option that doesn’t require touching a hyperscaler’s cloud instance pricing at all.
Server Partners: Dell, HPE, Supermicro, and Gigabyte Sign On
AMD didn’t launch the MI350P alone — it launched with a full bench of OEM server partners, which is itself a signal of how the company intends to sell this card: through the traditional enterprise server channel, not direct-to-hyperscaler.
- Dell Technologies confirmed MI350P support in its PowerEdge XE7745 and R7725 servers, with availability starting in July 2026.
- HPE listed support in its ProLiant DL380a Gen12 platform.
- Supermicro announced a 5U chassis capable of housing up to 10 MI350P cards in a single system.
- Gigabyte joined as a launch partner with its own server systems built around the card.
Notably absent from the launch materials: any of the major hyperscale cloud providers. No AWS, Azure, or Google Cloud commitment was announced alongside the MI350P, reinforcing that this specific SKU is aimed at the enterprise and channel market rather than hyperscale capacity — a different battlefield than the one AMD’s OAM parts fight on.
The Numbers Behind AMD’s Data Center Surge
The MI350P landed in the middle of AMD’s strongest data center quarter on record. AMD’s first-quarter 2026 financial results, reported May 5, showed total company revenue of $10.3 billion, up 38% year over year and ahead of Wall Street’s $9.9 billion consensus estimate. The Data Center segment — which bundles EPYC server CPUs with Instinct GPU shipments — brought in $5.8 billion, up 57% year over year from roughly $3.7 billion a year earlier.
AMD guided to roughly $11.2 billion in total revenue for the following quarter, and the market reaction to the print was immediate: AMD shares jumped as much as 18% in the days following the earnings report. None of that revenue is attributable to the MI350P specifically — the card wasn’t even announced until two days after the earnings call — but it establishes the financial backdrop the card is launching into: a data center business growing fast, but still tiny next to Nvidia’s.
Nvidia’s Scale Problem for AMD, By the Numbers
Here’s the number that puts the MI350P launch in perspective. Nvidia’s first-quarter fiscal 2027 results, covering the quarter ended April 26, 2026 and reported May 20, showed Data Center revenue of $75.2 billion — up 92% year over year and up 21% sequentially. Total Nvidia revenue for the quarter was $81.6 billion, up 85% year over year.
| Metric (most recent reported quarter, 2026) | AMD | Nvidia |
|---|---|---|
| Reporting period | Q1 2026 (ended ~March 2026) | Q1 FY2027 (ended April 26, 2026) |
| Total company revenue | $10.3 billion | $81.6 billion |
| Data center segment revenue | $5.8 billion | $75.2 billion |
| Data center YoY growth | +57% | +92% |
| Stock reaction to earnings | Shares up as much as 18% | Shares rose on the print |
Set side by side, Nvidia’s data center business alone generated roughly 13 times more revenue than AMD’s entire data center segment in a broadly comparable quarter — and Nvidia’s growth rate is still outpacing AMD’s. That’s the scale problem the MI350P doesn’t solve and was never going to solve on its own. What it can do is chip away at the deals Nvidia can’t fulfill because of allocation constraints, and give AMD a toehold in the enterprise inference market before that segment matures.
Historical Context: From MI210 to MI350P
AMD’s Instinct line has been chasing Nvidia’s data center GPUs generation after generation since well before the MI350P. The MI210, launched in 2022, was AMD’s last PCIe-format Instinct card before the company pivoted almost entirely to OAM modules for the MI300X (2023) and MI325X (2024) — both aimed squarely at hyperscaler deployments.
The MI325X comparison against Nvidia’s H200 previewed exactly the kind of claims AMD is now making with the MI350P. According to CRN’s reporting at the time, AMD said the MI325X “provides 40 percent faster throughput with an 8-group, 7-billion-parameter Mixtral model; 30 percent lower latency with a 7-billion-parameter Mixtral model and 20 percent lower latency with a 70-billion-parameter Llama 3.1 model” versus the H200. The MI350 series itself was first unveiled at AMD’s Advancing AI event in June 2025, with the MI350X and MI355X OAM variants shipping to hyperscalers well before the MI350P’s PCIe version arrived nearly a year later. The MI355X, the largest MI350-family part, packs 288GB of HBM3E and 8 TB/s of bandwidth — roughly double the MI350P’s memory bandwidth, reflecting how much AMD scales back its PCIe parts to fit standard power and cooling limits.
That multi-year pattern — matching Nvidia’s flagship on paper, one generation behind, at a PCIe or OAM format Nvidia has already established — is the throughline connecting the MI210 to the MI325X to the MI350P.
Industry Reactions
AMD’s own marketing frames the MI350P’s value proposition around total cost of ownership rather than pure benchmark supremacy. In its official positioning for the MI350 series, AMD says the goal is to “achieve leadership OPEX versus the NVIDIA H200 NVL card through lower precision MXFP6 and MXFP4 data types, more HBM3E memory more raw performance and an open low and no-cost software ecosystem for generative and agentic AI workloads,” according to AMD’s own Instinct MI350 product page.
Trade press coverage has largely framed the launch as AMD finally fielding a credible PCIe answer to Nvidia, rather than a knockout blow. Tom’s Hardware wrote plainly that “the introduction of the MI350P finally gives AMD a proper competitor to Nvidia’s fastest PCIe AI accelerator, currently the H200 NVL,” per its launch-day report. StorageReview’s coverage focused on the power-envelope match-up, noting that “the 600W rating also places the MI350P in the same bracket as NVIDIA’s H200 NVL and the RTX Pro 6000 Server, which it will be cross-shopped against in this segment,” in its write-up of the launch.
Competitive Landscape: More Than Just AMD vs Nvidia
The MI350P’s most direct fight is with Nvidia’s H200 NVL and, to a lesser extent, the memory-lighter RTX Pro 6000 Server Edition. But it’s launching into a market that’s gotten more crowded on multiple fronts at once. AMD isn’t the only company trying to chip into Nvidia’s data center dominance — Huawei’s homegrown AI chips are projected to generate roughly $12 billion in revenue as Chinese fabs race to keep pace with domestic AI model demand, while Arm has pushed roughly $2 billion in AI-capable CPU sales that still represent less than 5% of the broader data center compute market.
There’s also a newer threat neither AMD nor Nvidia fully controls: custom silicon built by the AI labs themselves. OpenAI’s Broadcom-fabbed “Jalapeño” accelerator, reportedly priced well below Nvidia’s equivalent parts, is part of a broader trend of hyperscalers and AI labs designing their own chips to reduce dependence on both AMD and Nvidia merchant silicon. That doesn’t directly threaten the MI350P’s enterprise/channel market today, but it’s a reminder that the AMD-vs-Nvidia framing, while the dominant story, isn’t the only one shaping where AI compute dollars go in 2026.
Market Impact: What This Means for Enterprise AI Buyers
For IT buyers, the MI350P’s real significance isn’t the benchmark table — it’s optionality. Two years of Nvidia H100 and H200 allocation constraints have left plenty of enterprises unable to get GPU capacity on their own timeline, regardless of budget. A PCIe card that drops into existing air-cooled racks, backed by four major server OEMs, gives procurement teams a second qualified vendor without a facilities overhaul.
That’s a meaningfully different pitch than AMD’s OAM parts make to hyperscalers, where the sales cycle is about winning a slice of a handful of massive capital budgets. The MI350P instead competes deal-by-deal against Nvidia’s H200 NVL and RTX Pro 6000 Server Edition in the enterprise and mid-market channel — a fragmented, higher-volume market where AMD’s ROCm software gap with Nvidia’s CUDA ecosystem has historically been the bigger obstacle to adoption than raw silicon specs. Whether the MI350P moves the needle depends less on the TFLOPS table and more on how far AMD has closed that software gap since the MI325X generation.
What’s Next: MI400, Helios, and AMD’s Roadmap
The MI350P is a bridge product, not AMD’s next big swing. AMD’s roadmap already points to the MI400-class accelerators and the rack-scale “Helios” systems built around them as the company’s next major step in hyperscaler-class AI infrastructure — the segment where the real dollars, and Nvidia’s real dominance, sit. Our earlier coverage of AMD’s Advancing AI roadmap and Helios rack bookings goes deeper on that side of the story.
The MI350P’s job in the meantime is narrower: keep AMD’s enterprise and channel business growing while the MI400 generation matures, and give ROCm more real-world deployments to harden against before AMD asks hyperscalers to bet even bigger on it.
Predictions: Where the MI350P Story Goes From Here
- Firm list pricing still hasn’t materialized. As of August 2026, AMD had not published official pricing even with Dell, HPE, and Supermicro systems past their planned July 2026 shipping window. Street pricing has settled in the $30,000-to-$40,000 range — comparable to Nvidia’s H200 NVL rather than a clear undercut — suggesting AMD is competing on memory capacity and compute specs rather than price, at least for now.
- Nvidia responds on price or bundling, not just specs. With the RTX Pro 6000 Server Edition already in the same segment, expect Nvidia to lean on aggressive channel pricing or software bundling rather than a pure spec refresh to defend its enterprise PCIe share through the rest of 2026.
- ROCm investment becomes the real story, not TFLOPS. AMD’s biggest obstacle in this segment has never been peak compute — it’s software maturity. Expect AMD to keep pouring marketing and engineering resources into ROCm compatibility claims through 2026 and 2027 to convert the MI350P’s spec advantages into actual enterprise design wins.
- MI350P sales stay a rounding error next to MI400/Helios. Given AMD’s own roadmap emphasis on MI400-class rack systems, expect the MI350P to matter more as a channel-building and ROCm-hardening exercise than as a meaningful line item in AMD’s data center revenue through 2026.
- The AMD-Nvidia scale gap persists through 2026. Even with AMD’s data center revenue growing 57% year over year, Nvidia’s data center base is roughly 13 times larger and still growing faster in percentage terms — meaning the gap in absolute dollars will likely widen before it narrows, regardless of how competitive any single AMD card looks on paper.
Frequently Asked Questions
What is the AMD Instinct MI350P?
The MI350P is AMD’s PCIe-format AI accelerator card, announced May 7, 2026, built on the CDNA 4 architecture with 144GB of HBM3E memory. It’s designed to run enterprise AI inference in standard air-cooled servers rather than the liquid-cooled racks AMD’s OAM-based Instinct cards require.
How much does the AMD MI350P cost?
AMD has not published an official list price, and as of August 2026 that remains unchanged. Market reporting since launch places real-world pricing in the $30,000-to-$40,000 range per card, comparable to street pricing for Nvidia’s H200 NVL rather than a clear discount to it.
Is the MI350P faster than Nvidia’s H200 NVL?
On AMD’s own theoretical peak-compute figures, yes: AMD claims roughly 20% better FP64, 43% better FP16, and 39% better FP8 performance than the H200 NVL. Nvidia’s H200 NVL still leads on memory bandwidth, at 4.8 TB/s versus the MI350P’s 4 TB/s. These are vendor-published theoretical peaks, not independently verified real-world benchmarks.
Which server makers support the MI350P?
Dell (PowerEdge XE7745 and R7725, slated for July 2026), HPE (ProLiant DL380a Gen12), Supermicro (a 5U chassis holding up to 10 cards), and Gigabyte all confirmed MI350P support at launch. No hyperscale cloud provider was named as a launch partner, and none of the four OEMs had published volume shipment figures as of August 2026.
How is the MI350P different from the MI350X and MI355X?
The MI350X and MI355X are OAM-format modules built for liquid-cooled hyperscaler racks; the MI355X carries 288GB of HBM3E and 8 TB/s of bandwidth, roughly double the MI350P’s specs. The MI350P is effectively a scaled-down, PCIe-compatible version of that same CDNA 4 family, built to fit a 600W air-cooled power envelope.
Can the MI350P run in any existing GPU server?
It’s built to the standard PCIe Gen 5 x16, full-height, full-length, dual-slot form factor, so it fits most modern air-cooled GPU server chassis. Actual support still depends on server vendor qualification — Dell, HPE, Supermicro, and Gigabyte have confirmed specific qualified platforms.
Does AMD sell the MI350P directly to cloud providers like AWS or Azure?
Not based on the launch materials. The MI350P is being sold through OEM server partners into the enterprise and channel market, not marketed as a hyperscale cloud SKU. AMD’s hyperscaler-facing push runs through its OAM-format MI300X, MI325X, MI350X, and MI355X lines instead.
What software does the MI350P run on?
AMD’s ROCm open-source software stack, which the company markets as a lower- or no-cost alternative to Nvidia’s CUDA ecosystem. ROCm maturity and third-party framework support remain the most commonly cited gap between AMD and Nvidia in enterprise AI deployments.
Related Coverage
- AMD Advancing AI: Helios Hits $5.25M, 12GW Booked [2026]
- Nvidia, AMD Hike GPU Kit Prices: China Retail Up 22% [2026]
- RTX 5090 vs RTX 5070 Ti for AI: 2x VRAM, 2.7x Price [2026]
- RAM Shortage 2026: AI Chips Hit 5 Gaming Platforms’ Prices
- Data Center Leases Hit $850B, Meta, Microsoft Lead [2026]
- OpenAI’s Jalapeño Chip Undercuts Nvidia by 50% [2026]
