On this page · 11 sections
- What changed in 2026, and why 2027 is still mostly roadmap
- What AMD has confirmed about MI500 and Helios 500
- What NVIDIA has said about Rubin Ultra, and what has been reported
- Head-to-head for buyers: memory, interconnect, software, availability
- The racks you can actually order for 2027
- Decision table: which way to lean in 2027
- Worked example: why the memory figure changes your rack count
- India, EU and US notes
- Checklist: what to get in writing before you commit
- How eCorpIT can help
- References
Summary. Neither 2027 rack can be bought on a spec sheet yet. AMD says Instinct MI500 GPUs launch in 2027 on its CDNA 6 architecture, a 2 nm process and HBM4E memory, and will power the Helios 500 rack with EPYC Verano CPUs (now named EPYC 9006 LP) and Pensando Como and Monza networking, followed by MI600 and Helios 600 in 2028. AMD has published no MI500 memory capacity, bandwidth, GPU count, power or price. NVIDIA's official target is Rubin Ultra systems in the second half of 2027, and its 2025 keynote slides showed an NVL576 system with 15 exaFLOPS of FP4 inference and 1 TB of HBM4e per GPU package. In July 2026 analyst firm SemiAnalysis reported that the Kyber rack had slipped to 2028; NVIDIA answered only that its roadmap was intact. The racks you can get now are Helios with 72 MI455X GPUs and 31 TB of HBM4, and Vera Rubin NVL72 with 72 Rubin GPUs and 20.7 TB. Our verdict: buy 2027 capacity on those, choose ROCm or CUDA on your own kernels rather than peak FLOPS, and write upgrade rights, slip clauses and memory guarantees into any multi-year contract.
About this article. For infrastructure, procurement and finance leads comparing AI rack roadmaps for 2027 and 2028. On 7 November 2026 we checked every date and specification against AMD and NVIDIA press releases, product pages, datasheets, keynote decks and earnings-call transcripts. Leaked or rumoured specifications are left out, reported delays are attributed, and NVIDIA's 2025 Rubin Ultra figures are labelled as 2025 targets. We have not tested this hardware, every performance claim is vendor-run, this is not financial advice, and no vendor paid for inclusion. Researched with AI assistance and reviewed by our editorial team.
What changed in 2026, and why 2027 is still mostly roadmap
AMD went from roadmap to shipping racks. AMD launched the Instinct MI400 Series and the Helios rack on 23 July 2026. On 3 November it reported record data centre segment revenue of $8.5 billion, up 97% year over year, said Anthropic, Meta, Microsoft, OpenAI and Oracle Cloud Infrastructure are running workloads on the first Helios racks, and said it expects growth to speed up in the fourth quarter as Helios and Venice production scales.
AMD named the 2027 rack and gave it a direction, not a size. At CES in January 2026, AMD said MI500 would use CDNA 6, a 2 nm process and HBM4E. The July launch named the Helios 500 parts, and the August earnings call promised larger scale-up domains over copper and optical interconnects. On 27 October AMD said its Meta work is informing MI500 and that the two are collaborating on Verano.
NVIDIA shipped Vera Rubin, and Rubin Ultra drew reports. NVIDIA said on 31 May that Vera Rubin was ramping into full production, told investors on 26 August that production shipments had begun that month, and announced general availability on Oracle Cloud Infrastructure on 27 October. Rubin Ultra remains a second-half-2027 target, questioned since June by reports on Kyber's timing, package and memory. NVIDIA has confirmed none of them.
What AMD has confirmed about MI500 and Helios 500
AMD's Financial Analyst Day on 11 November 2025 put MI500 on its roadmap for 2027, with a performance chart that carried no numbers. CES 2026, the July launch and keynote deck, and the August earnings call added the outline below.
| Item | What AMD has said | Status on 7 Nov 2026 |
|---|---|---|
| MI500 timing | 2027 (FAD 2025, CES 2026, Advancing AI 2026) | Announced roadmap; no month or quarter |
| Architecture and process | CDNA 6, 2 nm | Announced at CES 2026 |
| Memory | HBM4E | Type only; no capacity or bandwidth |
| Performance claim | More than 2,000x MI300X inference throughput in four years | AMD engineering projection |
| Scale-up | Larger scale-up domain, copper and optical interconnect | No domain size published |
| Helios 500 host CPU | EPYC Verano, now named EPYC 9006 LP, LPDDR memory, up to 5 GHz | Announced; Meta collaborating on Verano |
| Helios 500 networking | Pensando Como and Monza | Names only |
| Rack power, cooling, price | Not stated | Not published |
| Successor | MI600 and Helios 600 in 2028, with EPYC Ferrara and Pensando Palma and Levanzo | Announced roadmap |
Two cautions. AMD's endnote bases the 2,000x figure on projected peak performance of an MI500-powered rack against an MI300X platform, so it says nothing about MI500 against Rubin Ultra. And AMD's product page still calls today's Helios a reference design for OEM and ODM partners, not an AMD product for sale; systems come from Bull, HPE, Lenovo, Supermicro, Sanmina and Wiwynn, and HPE's first go to Vultr. Expect Helios 500 quotes, warranties and support through the same channel.
What NVIDIA has said about Rubin Ultra, and what has been reported
NVIDIA's GTC 2025 blog put Rubin Ultra systems in the second half of 2027; the detailed figures come from that keynote's slides, as reported by Tom's Hardware. In October 2025 NVIDIA said Kyber, the successor to its Oberon rack, would house 576 Rubin Ultra GPUs by 2027. At GTC 2026 it demonstrated a Rubin Ultra tray, and Tom's Hardware described Kyber as holding 144 GPU packages in vertical trays, liquid-cooled by default.
| Claim | Source | How firm |
|---|---|---|
| Rubin Ultra systems in 2H 2027; Kyber to house 576 GPUs | NVIDIA blogs, 2025 | Official target |
| NVL576: 15 EF FP4 inference, 5 EF FP8 training, 365 TB fast memory | GTC 2025 keynote slide, per Tom's Hardware | 2025 target |
| Per package: four GPU dies, 1 TB HBM4e, 100 PFLOPS FP4 | GTC 2025 slide; tray shown again at GTC 2026 | 2025 target |
| NVLink 7 at 1.5 PB/s; CX9 at 115.2 TB/s per NVL576 | GTC 2025 keynote slide, per Tom's Hardware | 2025 target |
| Full-scale 800 VDC production alongside Kyber in 2027 | NVIDIA Technical Blog, May 2025 | Official target |
| Kyber delayed by more than 12 months, into 2028 | SemiAnalysis, per Tom's Hardware, 6 July 2026 | Report; NVIDIA says roadmap intact |
NVIDIA's one-sentence reply did not say whether the intact roadmap was the original schedule or a revised one. Tom's Hardware, summarising SemiAnalysis, said the largest Rubin Ultra scale-up domain in 2027 could equal, but not beat, today's 72-package Oberon rack. Other reports since June say NVIDIA is reworking the package and memory. We leave their figures out because NVIDIA has not confirmed them; it says only that it is optimising compute, networking and memory so it can build more GPUs. Our GTC 2027 buyer briefing covers what to ask in March, and our Rubin Ultra NVL576 Kyber rack guide covers power and readiness.
Head-to-head for buyers: memory, interconnect, software, availability
What each vendor has put on record for 2027, from the sources above. Thin AMD cells mean AMD has published little, not that the product is weaker.
| Dimension | AMD MI500 and Helios 500 | NVIDIA Rubin Ultra | Label |
|---|---|---|---|
| Timing | 2027 | Systems 2H 2027; Kyber delay to 2028 reported | Roadmap |
| HBM per GPU package | HBM4E; capacity not published | 1 TB HBM4e (2025 target) | Unconfirmed |
| Low-precision compute | Not published | 15 EF FP4 per NVL576 (2025 target) | Unconfirmed |
| Scale-up fabric | Larger domain over copper and optics; size not published | NVLink 7, 1.5 PB/s across NVL576 (2025 target) | Roadmap |
| Scale-out networking | Pensando Como and Monza named | ConnectX-9, 115.2 TB/s per NVL576 (2025 target) | Roadmap |
| Software | ROCm, open source; ROCm 10 released 27 Aug 2026 | CUDA and NVIDIA libraries | Shipping today |
| Rack design | Not published | Kyber, or today's NVL72 rack if Kyber slips | Roadmap |
| Power | Not published | 800 VDC, racks of 1 MW and beyond | NVIDIA target |
Memory: AMD's lead today, nobody's guarantee for 2027
Memory is the one current comparison that holds up without normalisation: up to 432 GB of HBM4 per MI455X and 31 TB per Helios rack, against 288 GB per Rubin GPU and 20.7 TB per Vera Rubin NVL72. For 2027, AMD has named HBM4E but no capacity, and NVIDIA's 1 TB per package is a 2025 figure on a design it says it is still optimising.
Supply is the constraint. NVIDIA's chief financial officer told investors on 26 August that memory prices had beaten NVIDIA's expectations and would climb into 2027, and that supply would stay a bottleneck at least to the end of fiscal 2028, which runs to late January 2028. AMD lists memory supply among its risks. Ask both vendors the same memory question and put the answer in the contract; our HBM4E supply and pricing guide covers the market side.
Interconnect: open standards against a proprietary fabric
Today's Helios links 72 GPUs with UALink over Ethernet at up to 260 TB/s of aggregate scale-up bandwidth, plus 43 TB/s of Ultra Ethernet-aligned scale-out. Vera Rubin NVL72 lists 216 TB/s of NVLink 6 and 32.4 TB/s of networking, which NVIDIA marks as bi-directional. AMD does not state directionality, which can hide a 2x difference; our Helios vs Vera Rubin spec-trap guide explains how to normalise the two.
For 2027 the question is scale-up domain size. NVIDIA's 2025 answer was 576 GPUs; the reported Kyber delay points to 72 packages in practice. AMD promises a larger MI500 domain, without a number. Put domain size, link bandwidth and direction into the request for proposal.
Software: ROCm vs CUDA is an engineering estimate, not a belief
AMD released ROCm 10 on 27 August 2026 with ROCm.AI generally available: Hyperloom, an agent that optimises inference workloads on vLLM and SGLang, plus AMD Skills for coding agents such as Claude Code, Codex and Cursor. AMD claims an average 3.3x inference and 2.4x training gain over ROCm 7 on the same MI355X systems; that is vendor-run and compares AMD releases, not AMD against NVIDIA. AMD says PyTorch, Hugging Face, vLLM and SGLang already run on MI455X, and that OpenAI is optimising GPT-class workloads with its Triton framework on ROCm.
CUDA remains the incumbent, and the switching cost sits in custom kernels, tuned inference servers, profilers and team habits, not framework support. A stack built mostly on PyTorch, vLLM or SGLang ports more easily; hundreds of hand-written CUDA kernels need pricing before any hardware discount. Our ROCm vs CUDA lock-in guide sets out how to inventory that risk.
Availability and cloud partners
AMD names OpenAI, Anthropic, Meta, Microsoft, Oracle, HUMAIN, Tensorwave, Vultr and Cirrascale among Helios customers. OpenAI expects Helios online from the fourth quarter of 2026, with deployments speeding up through 2027, and Anthropic plans up to 2 GW of MI455X GPUs in Helios, the first gigawatt from the first half of 2027. Vera Rubin racks run at CoreWeave, Google Cloud, Microsoft Azure, Oracle Cloud Infrastructure and Nebius, and NVIDIA and Oracle announced a multi-gigawatt Vera Rubin expansion on 27 October 2026. We found no announcement of MI500 or Rubin Ultra cloud instances as of 7 November 2026.
The racks you can actually order for 2027
Most 2027 budgets will buy or rent the current generation. These figures come from AMD's Helios page and launch release and NVIDIA's product page and October 2026 datasheet, each on that vendor's own terms.
| Specification | AMD Helios (MI455X) | NVIDIA Vera Rubin NVL72 |
|---|---|---|
| Accelerators and CPUs | 72 MI455X, 18 EPYC Venice | 72 Rubin GPUs, 36 Vera CPUs |
| HBM per GPU | Up to 432 GB HBM4 | 288 GB HBM4 |
| HBM per rack | 31 TB | 20.7 TB |
| Memory bandwidth per GPU | Up to 23.3 TB/s (19.6 TB/s in the same page's tray specification) | 19.2 TB/s |
| Low-precision compute per rack | 2.9 EF FP4, 1.4 EF FP8 (peak) | 3,600 PF NVFP4 sparse; 2,520 PF NVFP4 and 1,260 PF FP8 dense |
| Scale-up | 260 TB/s, UALink over Ethernet | 216 TB/s, NVLink 6 |
| Scale-out | 43 TB/s | 32.4 TB/s, bi-directional |
| Host CPU cores | Venice, up to 256 cores each | Vera, 88 cores each (3,168 per rack) |
| Rack power | No AMD figure found | 227 kW MaxP, 187 kW MaxQ, 156 kW MaxLPS |
| Availability | First racks running at five customers; volume expected 2H 2026 | Shipping since August 2026; generally available on OCI |
AMD claims up to 30% more tokens per dollar than Vera Rubin NVL72, an AMD Performance Labs estimate on Kimi K2 Thinking at 32K input and 8K output tokens using AMD's own projection of hourly pricing. NVIDIA claims one-tenth the cost per million tokens of GB200 NVL72 on the same model and sequence lengths. Both are vendor-run: reasons to benchmark, not results.
Decision table: which way to lean in 2027
| Your situation | Lean towards | Why |
|---|---|---|
| Need capacity in the first half of 2027 | Helios (MI455X) or Vera Rubin NVL72 now | MI500 and Rubin Ultra have no confirmed ship dates |
| Memory-bound inference: long context, large mixture-of-experts | Shortlist Helios; ask for MI500 memory in writing | 31 TB vs 20.7 TB per rack is AMD's clearest published lead |
| Large estate of hand-written CUDA kernels | Vera Rubin now, Rubin Ultra once dated | Porting cost can outweigh any hardware discount |
| Mostly PyTorch, vLLM or SGLang, few custom kernels | Run a Helios pilot before committing | Smaller port; ROCm 10 targets these frameworks |
| Want open fabrics and several rack suppliers | Helios family | UALink, Ultra Ethernet and OCP Open Rack Wide |
| Building a hall that opens in 2028 or later | Design for liquid cooling and an 800 VDC path | NVIDIA designs for 1 MW racks; AMD has published no Helios 500 power |
| Signing a multi-year commitment | Either, with clauses | Upgrade rights, slip penalties and memory guarantees protect you |
Worked example: why the memory figure changes your rack count
Take one serving replica of a 2-trillion-parameter model at 4-bit precision. Weights alone are 2 × 10^12 × 0.5 bytes = 1 TB. Add 1 TB for key-value cache and activations at long context, our assumption, and each replica needs about 2,000 GB of GPU memory.
Minimum accelerators per replica, by capacity only:
- MI455X at 432 GB: 2,000 ÷ 432 = 4.6, so 5 GPUs.
- Rubin at 288 GB: 2,000 ÷ 288 = 6.9, so 7 GPUs.
- Rubin Ultra at the 2025 target of 1 TB per package: 2,000 ÷ 1,000 = 2 packages, with no headroom.
- Rubin Ultra if a shipped package carried 25% less memory, 750 GB: 2,000 ÷ 750 = 2.7, so 3 packages. This is our illustration, not an NVIDIA figure.
- MI500: cannot be calculated, because AMD has published no capacity.
Replicas per 72-accelerator rack: Helios fits 72 ÷ 5 = 14 replicas and Vera Rubin NVL72 fits 72 ÷ 7 = 10, each with two GPUs spare. Pooled capacity agrees: 72 × 432 GB = 31,104 GB against 72 × 288 GB = 20,736 GB.
Parallelism layouts round these numbers up: if your stack shards in groups of 8, both racks hold 72 ÷ 8 = 9 replicas and AMD's capacity lead disappears for this model. Throughput also depends on bandwidth and software. The point is sensitivity: a 25% cut in memory per package raises Rubin Ultra packages per replica by half, which flows into rack count, power and price. The memory configuration belongs in the contract.
India, EU and US notes
India. On 16 February 2026, AMD and TCS said TCS subsidiary HyperVault AI Data Center will co-develop a Helios-based rack-scale design supporting up to 200 MW in India, which TCS presented as the start of Helios-powered AI infrastructure in the country. Neither has published rupee pricing or a go-live date. Domestic Vera Rubin capacity is ordered but not yet live at scale, as our GTC briefing sets out. For personal data under the DPDP Act, confirm where capacity physically sits before you sign.
EU. AMD says the EuroHPC Joint Undertaking selected MI430X GPUs and 6th Gen EPYC CPUs for the LUMI-AI supercomputer. Neither vendor has announced European MI500 or Rubin Ultra capacity.
US. NVIDIA's May 2025 post set out a move to 800 VDC power for 1 MW racks from 2027, with Eaton, Schneider Electric and Vertiv among its partners, and tied full-scale production to Kyber. Ask colocation providers for a phased conversion path rather than a date; our 800 VDC power guide covers the options.
Checklist: what to get in writing before you commit
- HBM capacity, stack count and generation per package, with a remedy if the shipped part has less.
- Compute dies per package, and the datatype and sparsity behind every FLOPS figure.
- Scale-up domain size, link bandwidth and direction, and the rack design (Kyber, today's NVL72 rack or Helios 500).
- Sustained rack power in kW, coolant inlet temperature and any 800 VDC need.
- A committed delivery quarter, a slip penalty and a right to substitute current-generation racks.
- Upgrade rights from Helios or Vera Rubin positions to MI500 or Rubin Ultra.
- Sustained tokens per second on your model, batch size and context length.
- The legal entity carrying the warranty, which for Helios is the OEM, not AMD.
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 help infrastructure buyers turn vendor roadmaps into dated plans, from rack power and cooling assessments to colocation and GPU cloud evaluation, through our data centre and cloud networking services. Talk to us at /contact-us/.
Last updated: 7 November 2026.
