On this page · 13 sections
- What AMD launched on 10 November 2026
- The SP7 lineup: nine models, all with 1Ku prices
- Platform: memory, I/O and rack power
- Venice vs Xeon 6 vs NVIDIA Vera: specs and prices
- Reading the benchmark claims, all vendor-run
- What Venice means for a 2026 to 2027 server refresh
- Decision table: which CPU to choose
- Worked example: an N+1 Hyper-V cluster for a 1,200-employee US company
- Cloud options: Azure HDv2, DHasv8 and OCI E7
- Notes for US and UK buyers
- Implementation checklist
- How eCorpIT can help
- References
Summary. On 10 November 2026 AMD announced general availability of its 6th Gen EPYC SP7 server processors, codenamed Venice and sold as the EPYC 9006 Series, and said systems are shipping from leading server makers. AMD's specification list shows nine SP7 models from 64 to 256 cores, each with a 1Ku list price, from $8,008 to $14,904. The flagship EPYC 9996 has 256 cores, 512 threads, a 4.1 GHz maximum boost, 1,024 MB of L3 cache and 600 W of Default CPU Power for $14,904 per unit in 1,000-unit quantities. Every SP7 part has 16 DDR5 channels, MRDIMM support at 12,800 MT/s for up to 1.6 TB/s per socket, and PCIe 6.0. Intel's 128-core Xeon 6980P lists at $13,955 with 12 channels and PCIe 5.0. NVIDIA publishes no price for its 88-core Vera. AMD claims a two-socket 9996 system delivers 2.2x a Xeon 6980P system on SPECrate2017_int_base and 3.7x a Vera system, but the Vera figure is an AMD estimate. Azure HDv2 is in preview and OCI E7 is coming, without public prices. Our verdict: Venice is the default for consolidation, VM density and agent sandboxes if your racks can feed 500 W to 600 W sockets. Per-core licences and N+1 host maths, not the headline multiplier, decide the case against Xeon 6.
About this article. For infrastructure owners and finance leads in the US and UK planning a 2026 or 2027 server refresh. On 21 November 2026 we checked launch facts, specifications and prices against AMD's investor relations release, launch blog, newsroom and specification pages, Intel's and NVIDIA's product pages, published SPEC CPU results, Microsoft's pricing page and Azure update feed, Broadcom's licensing note, GOV.UK and the European Central Bank. AMD's website refused automated access, so we read archived copies of its specification list (4 November 2026) and EPYC 9996 page (18 November 2026). We have not benchmarked these processors: every performance figure is vendor-run, vendor-submitted or vendor-estimated. This is not licensing, tax or financial advice, vendors change prices, and no vendor paid for inclusion. Researched with AI assistance and checked against the primary sources listed in the references.
What AMD launched on 10 November 2026
AMD's release says 6th Gen EPYC SP7 systems are shipping from leading server vendors and that Microsoft Azure and Oracle Cloud Infrastructure instances will be generally available soon. Three further details matter to buyers.
- SP7 is the first of four families. AMD's launch blog says EPYC 9006 SP8, 9006X SP7 and 9006 LP processors will follow, without dates. AMD positions SP8 for general enterprise work, 9006X for HPC and technical computing with extra cache, and LP for AI host nodes. Its specification list already shows 22 SP8 models from 8 to 128 cores, with eight memory channels and 1Ku prices from $700 to $11,679.
- Named server makers. The blog quotes Dell, HPE (with four new ProLiant Gen13 servers), Lenovo and Supermicro. AMD's press kit links partner announcements from Cornelis, Hyve Solutions, Lenovo, MSI, QCT and Supermicro, whose release headline says it is now shipping.
- A lead customer. AMD says Meta is validating 6th Gen EPYC platforms as a lead customer ahead of deployment at scale.
Our July analysis of Venice and the 2027 refresh call predates general availability, and our Diamond Rapids vs EPYC comparison covers Intel's 2027 roadmap. This article is the launch-facts reference.
The SP7 lineup: nine models, all with 1Ku prices
The table comes from AMD's server processor specification list as archived on 4 November 2026, and the EPYC 9996 row also matches AMD's 9996 product page as archived on 18 November. Default CPU Power is AMD's new term in place of TDP; the bracket shows the configurable range.
| Model | Cores / threads | Max boost | Base clock | L3 cache | Default CPU Power (range) | 1Ku price |
|---|---|---|---|---|---|---|
| EPYC 9996 | 256 / 512 | 4.1 GHz | 2.55 GHz | 1,024 MB | 600 W (400 to 600 W) | $14,904 |
| EPYC 9966 | 192 / 384 | 4.0 GHz | 2.9 GHz | 768 MB | 600 W (400 to 600 W) | $14,079 |
| EPYC 9846 | 168 / 336 | 3.7 GHz | 2.85 GHz | 768 MB | 500 W (320 to 500 W) | $13,114 |
| EPYC 9756 | 128 / 256 | 4.0 GHz | 3.15 GHz | 512 MB | 500 W (320 to 500 W) | $12,498 |
| EPYC 9G76 | 96 / 192 | 4.8 GHz | 3.4 GHz | 384 MB | 500 W (320 to 500 W) | $11,622 |
| EPYC 9686F | 96 / 192 | 5.0 GHz | 3.4 GHz | 384 MB | 500 W (400 to 600 W) | $11,434 |
| EPYC 9656 | 96 / 192 | 3.7 GHz | 3.05 GHz | 512 MB | 400 W (220 to 400 W) | $9,713 |
| EPYC 9586F | 64 / 128 | 5.0 GHz | 3.75 GHz | 384 MB | 500 W (400 to 600 W) | $9,701 |
| EPYC 9556 | 64 / 128 | 4.3 GHz | 2.75 GHz | 384 MB | 300 W (220 to 300 W) | $8,008 |
Four things stand out.
Prices are published. AMD's launch footnotes state CPU prices as of 5 November 2026 and give $14,904 for the 9996 and $11,988 for the previous-generation 192-core EPYC 9965. The same 1Ku figures sit against every model in AMD's list. A 1Ku price is per unit in 1,000-unit orders: list guidance, not what an OEM charges inside a configured server. Per core, the 9996 lists at about $58 ($14,904 / 256 = $58.22) against $109 for the 6980P ($13,955 / 128 = $109.02).
Check the final clock speeds. AMD's release and blog say high-frequency parts reach up to 5.3 GHz with 96 cores, and a blog footnote names the 9686F. AMD's specification list shows 5.0 GHz for that part, and the published SPEC CPU2026 result for a Dell PowerEdge R9825 with two 9686F processors records a 5,000 MHz maximum and a 3,400 MHz base. Ask your OEM which figure applies to the parts it ships.
Export controls. AMD's 9996 page labels the model as export controlled under ECCN 5A992.z.2 of the US Export Administration Regulations. We could not read current pages for the other models, so confirm the status of each ordering part number early if you ship servers across borders or buy through a distributor outside the US.
Two core types, configurable power. AMD combines Zen 6 cores for frequency with Zen 6c cores for density. The 9996 can be configured from 400 W to 600 W, and AMD's agents-per-watt footnote models a 256-core part at 400 W, which suits power-capped racks.
Platform: memory, I/O and rack power
Every SP7 model has 16 DDR5 channels, RDIMM support up to 8,000 MT/s (1,024 GB/s per socket) and MRDIMM support up to 12,800 MT/s (1,638 GB/s per socket). AMD's footnote compares 1,638.4 GB/s with 844.8 GB/s for Xeon 6 and Arm AGI, which use 12 channels of MRDIMM-8800. The arithmetic holds: 16 × 12,800 MT/s × 8 bytes = 1,638.4 GB/s, and 12 × 8,800 × 8 = 844.8 GB/s.
AMD's lane figures need care. Its specification list gives each SP7 model 96 PCIe 6.0 lanes, while a launch footnote says 9006 CPUs support up to 128 Gen 6 lanes in a single socket, the figure its SP8 listings show. AMD lists CXL 3.1 for the embedded SP7 parts launched the same day. Confirm usable lanes and CXL support on the exact motherboard before you plan GPU or NVMe counts.
Memory may cost more than the CPU. Sixteen channels per socket means 32 DIMMs per two-socket server at one DIMM per channel, against 24 on a 12-channel Xeon 6 platform. With DRAM prices high, a fully populated Venice server can carry a memory bill well above its CPU bill. Our 2027 server memory supply analysis covers pricing and allocation.
Rack power sets the server count. AMD's agent sandbox example uses 9996 servers with 16 DDR5-8000 RDIMMs per socket at 830 W of combined SoC and DIMM power per socket. In a 35 kW rack that fits 16 servers at default settings, or an estimated 18.8 if AMD's power-management feature caps combined power 15% lower (706 W per socket), at an estimated 1% performance cost based on testing of a 5th Gen part. That is AMD analysis, not a measurement, but the arithmetic transfers: 35,000 W / 16 = 2,187.5 W per server, so a 15 kW rack holds six such servers (15,000 / 2,187.5 = 6.9, rounded down). In a two-socket server, two sockets at 830 W use 1.66 kW before drives and NICs.
Venice vs Xeon 6 vs NVIDIA Vera: specs and prices
The EPYC figures come from AMD's specification pages and launch footnotes, the 9965 boost from its published SPEC result, the Intel figures from Intel's 6980P page, and the Vera figures from NVIDIA's product page, with clock and power taken from AMD's footnotes. The SPECrate scores are results published at spec.org, so they are vendor-submitted rather than independent tests.
| Item | EPYC 9996 (6th Gen) | EPYC 9965 (5th Gen) | Intel Xeon 6980P | NVIDIA Vera |
|---|---|---|---|---|
| Cores / threads per socket | 256 / 512 | 192 / 384 | 128 / 256 | 88 / 176 |
| Max boost | 4.1 GHz | 3.7 GHz | 3.9 GHz | Up to 3.2 GHz (per AMD) |
| Power | 600 W Default CPU Power | 500 W TDP | 500 W TDP | 450 W (per AMD) |
| Memory | 16 ch DDR5, MRDIMM-12800, up to 1.6 TB/s | Not compared here | 12 ch DDR5-6400 or MRDIMM-8800, max 3 TB | LPDDR5X SOCAMM, up to 1.2 TB/s, up to 1.5 TB |
| I/O | PCIe 6.0, 96 lanes listed | Not compared here | PCIe 5.0, 96 lanes | NVLink-C2C up to 1.8 TB/s to GPUs |
| List price per CPU | $14,904 (1Ku) | $11,988 (1Ku) | $13,955 (Intel RCP) | Not published |
| 2P SPECrate2017_int_base | 5,470 (published) | 3,260 (published) | 2,510 (published) | 1,459 (AMD estimate) |
| Status, 21 November 2026 | Shipping from OEMs | Current 5th Gen part | Launched Q3 2024 | Generally available on OCI |
On SPECrate2017_int_base per core, the 9996 scores 10.68 against 9.80 for the 6980P, about 9% higher (10.68 / 9.80 = 1.09). Most of the 2.2x system-level gain comes from having twice the cores per socket. That matters for software licensed per core, as the worked example shows.
Reading the benchmark claims, all vendor-run
The AMD claims below come from its release and blog footnotes, and the NVIDIA claims from its Vera page.
| Claim | Who makes it | Basis | How to read it |
|---|---|---|---|
| 9996 system has 2.2x the performance of a 6980P system | AMD | Published SPECrate2017_int_base results, 5,470 vs 2,510 | Real published scores; doubled core count does most of the work |
| 2.0x performance per dollar vs 6980P | AMD | Score divided by CPU list price | CPU prices only; excludes memory, servers and per-core software |
| 3.7x the performance of a Vera system | AMD | AMD estimate of 1,459 for 2P Vera, based on Phoronix testing | NVIDIA has published no SPECrate2017 score; unverified |
| About 3.3x agents per CPU watt vs Vera, 2.5x vs 6980P | AMD | Thread count used as a proxy for agents, 256 cores at 400 W | A density ratio, not measured agent throughput |
| 1.26x per-core SPECrate2026 vs Vera (EPYC 9686F) | AMD | Published 9686F result vs NVIDIA's own estimate for Vera | Compares a published score with an estimate |
| Up to 1.8x faster agentic sandbox work than leading x86 | NVIDIA | NVIDIA measured data on compiling, Python tool chains and code analysis | Baseline is the latest-generation x86 CPU, not an EPYC 9006 test |
| 80% faster agentic task completion than x86 | NVIDIA | NVIDIA's Vera page | No public workload detail on the page |
The fair reading: AMD leads on throughput per socket and density in published SPEC results, while NVIDIA argues Vera leads on single-thread performance and memory bandwidth per core for agent loops. Neither has published a head-to-head test on the same agent workload, so run your own harness on Vera on OCI and on Venice once a cloud preview or test server is available.
What Venice means for a 2026 to 2027 server refresh
Against Xeon 6. The 6980P is a mature part at a known price, and Intel lists accelerators such as AMX and QuickAssist on its specification page. Venice offers twice the cores per socket, 16 memory channels against 12 and PCIe 6.0 against PCIe 5.0. For general virtualisation and container hosts, Venice is the stronger platform on paper. For stacks validated only on Intel, or that rely on Intel accelerators, Xeon 6 remains a sound buy.
Against NVIDIA Vera. Vera is an Arm-compatible CPU with LPDDR5X SOCAMM memory and NVLink-C2C to NVIDIA GPUs. It became generally available on OCI on 27 October 2026, which NVIDIA repeated in its third-quarter results, as our NVIDIA earnings analysis covers. NVIDIA also sells it as a liquid-cooled CPU rack or in standard dual- and single-socket configurations. Vera makes most sense where agent execution sits next to NVIDIA GPU clusters and your software already runs on Arm. For a general-purpose x86 estate, Venice avoids a recompile and keeps existing hypervisors and licences.
Against waiting. SP8, 9006X and LP are announced but undated. Wait for SP8 only if your racks cannot feed 500 W to 600 W sockets or you need a smaller socket for edge sites; AMD's SP8 listings top out at 400 W.
Licensing decides more than silicon. Windows Server 2025 Datacenter has a Microsoft suggested price of $6,771 per 16 core licences. Broadcom licenses VMware Cloud Foundation and vSphere Foundation on every physical core, with a minimum of 16 cores per CPU. Covering a 512-core two-socket host with licences is expensive, and an N+1 spare host of that size carries the same licence bill as a production one.
Decision table: which CPU to choose
| Choose | If | Watch out for |
|---|---|---|
| EPYC 9996 or 9966 | You are consolidating many older hosts, running dense VMs or containers, or hosting agent sandboxes, and racks can supply the power | 384 to 512 cores per host enlarge failure domains and the N+1 licence bill |
| EPYC 9756 or 9846 | You want Venice memory and PCIe 6.0 with fewer cores per licensed host | No published SPECrate result for either yet |
| EPYC 9686F, 9G76, 9586F or 9556 | You need GPU host nodes or high per-core speed for per-core licensed databases | The F parts default to 500 W even at 64 or 96 cores |
| Intel Xeon 6980P | Your stack is validated on Intel, uses AMX or QAT, or you want a mature platform now | 12 memory channels and PCIe 5.0 |
| NVIDIA Vera | Agent execution runs beside NVIDIA GPUs on OCI and your code runs on Arm | No published price; vendor claims not independently tested |
| Wait for SP8, 9006X or LP | Racks are power-capped, or you need a cache-heavy or host-node part | No dates announced |
Worked example: an N+1 Hyper-V cluster for a 1,200-employee US company
A 1,200-employee US logistics company wants to replace an old Windows Server Hyper-V estate with a cluster that delivers about 20,000 SPECrate2017_int_base of throughput, plus one spare host for maintenance and failure (N+1). We use SPECrate only as a throughput proxy, the published scores and CPU list prices above, and Microsoft's $6,771 suggested price for Windows Server 2025 Datacenter per 16 core licences. We assume every physical core in each host is licensed; confirm your terms with your reseller.
| Line | Option A: 2P EPYC 9996 hosts | Option B: 2P Xeon 6980P hosts |
|---|---|---|
| Hosts to reach about 20,000 | 4 (4 × 5,470 = 21,880) | 8 (8 × 2,510 = 20,080) |
| Hosts including one spare | 5 | 9 |
| CPUs | 10 | 18 |
| CPU list cost | 10 × $14,904 = $149,040 | 18 × $13,955 = $251,190 |
| Physical cores needing licences | 5 × 512 = 2,560 | 9 × 256 = 2,304 |
| Windows Server 2025 Datacenter | 160 packs × $6,771 = $1,083,360 | 144 packs × $6,771 = $975,024 |
| CPU plus Windows subtotal (USD) | $1,232,400 | $1,226,214 |
| CPU plus Windows subtotal (GBP, before VAT) | About £936,600 | About £931,900 |
| CPU power at default or TDP | 10 × 600 W = 6.0 kW | 18 × 500 W = 9.0 kW |
| DIMM slots at one per channel | 160 | 216 |
The GBP row uses the European Central Bank reference rates for 20 November 2026 (USD 1.0990 and GBP 0.83735 per euro), crossed to £0.76 per US dollar and rounded to two decimals ($1,232,400 × 0.76 = £936,624). AMD and Intel publish these prices in US dollars.
What the numbers show:
- The CPU saving disappears into licences. Venice saves $102,150 on CPUs, but its 512-core spare host adds 256 more licensed cores than Intel's spare. The subtotals differ by $6,186, about 0.5%.
- Venice still wins on everything else. Four fewer servers means fewer chassis, NICs, switch ports and rack units, 3.0 kW less CPU power and 56 fewer DIMMs. Those lines are not priced here because they depend on your OEM and memory quotes, but at 2026 memory prices they are large.
- Scale changes the answer. Without a spare host, Option A costs $985,920 against $1,089,968 for Option B. At a larger 8+1 versus 16+1 design, Venice's CPU plus Windows subtotal is about $98,000 lower ($2,218,320 against $2,316,182). The bigger the cluster, the smaller the share of the N+1 licence penalty.
- A mid-core Venice part may be the real winner. A 128-core EPYC 9756 host matches Intel's core count per host at a 1Ku price of $12,498, $1,457 below the 6980P, while keeping 16 memory channels and PCIe 6.0. We found no published SPECrate result for it, so we have not sized it here; ask your OEM for one.
For UK buyers, the standard VAT rate is 20%. A VAT-registered business can normally recover VAT on business purchases, but confirm the treatment with your adviser.
Cloud options: Azure HDv2, DHasv8 and OCI E7
- Azure HDv2 is in preview for AI data processing, agentic AI and HPC, AMD says. AMD also names Azure HXv2 for EDA and technical computing and Azure ND MI455X v7, based on AMD's Helios rack, for production AI inference, without dates. Microsoft's quote in AMD's blog says customers can request HDv2 access and calls it the first 6th Gen EPYC platform in the cloud. We found no prices or GA dates.
- Azure DHasv8, DHalsv8 and EHasv8. On 17 November 2026 Microsoft announced a private preview of general-purpose and memory-optimised VMs on Venice, with sizes up to 480 cores and up to 3.8 TiB of memory on EHasv8. Its update lists no GA date.
- OCI Compute E7 instances will be powered by 6th Gen EPYC, AMD said on 27 October, for agent orchestration, concurrent agent sandboxes, retrieval services, in-memory caches and web and API tiers. E7 Acceleron Standard VMs will offer flexible compute and memory configurations.
Cloud-first teams should request preview access if agent or data-processing work justifies it, and keep pricing current instances, because neither provider has dated Venice GA.
Notes for US and UK buyers
United States. AMD's 1Ku figures and Intel's Recommended Customer Price are list guidance, not what an OEM charges inside a configured server. The 9996 is classified ECCN 5A992.z.2, which matters if you ship servers abroad.
United Kingdom. Budget in GBP with a currency buffer, because both CPU vendors price in US dollars. At £0.76 per dollar, a $14,904 EPYC 9996 is about £11,330 and a $13,955 Xeon 6980P about £10,610, before VAT. Ask your distributor whether export-controlled Venice models need extra end-user paperwork for your use.
Implementation checklist
- Record your contracted kW per rack and cooling type before choosing a SKU; 600 W sockets change server counts.
- Count licensed cores per host, including spare hosts, for Windows Server, VMware and any per-core database.
- Get OEM quotes for at least one 256-core and one 96- or 128-core Venice configuration, plus a Xeon 6 configuration.
- Price memory separately: 32 DIMMs per Venice server against 24 per Xeon 6 server at one DIMM per channel.
- Confirm PCIe 6.0 lane counts, CXL support and final clock speeds on the exact motherboard and part.
- Check whether your chosen model is export controlled and whether your distributor needs end-user details.
- For agent sandboxes, benchmark your own harness on Vera (OCI) and on Venice (an Azure preview or OEM test server).
- Revisit when AMD dates SP8, 9006X and LP, or when Azure or OCI publish Venice prices.
How eCorpIT can help
eCorpIT is a Gurugram-based technology organisation, founded in 2021, assessed at CMMI Level 5 and MSME certified, with senior-led engineering teams working across AWS, Microsoft and Google platforms. We turn launch claims into refresh plans, with licensed-core and N+1 modelling, rack power checks and the networking behind new hosts, through our data centre and cloud networking services. Talk to us at /contact-us/.
Last updated: 21 November 2026.