On this page · 12 sections
- What NVIDIA has confirmed about GTC 2027
- The roadmap NVIDIA has published
- Vera Rubin NVL72: the rack you can buy now
- Rubin Ultra and Kyber: promised versus reported
- 800 VDC and liquid cooling: decisions that outlast the chip
- Supply and lead times
- Questions to ask before and at GTC 2027
- Buy now, rent, or wait: a decision table
- Worked example: how many Rubin GPUs fit in a 10 MW hall
- India, EU and US notes
- How eCorpIT can help
- References
Summary. NVIDIA GTC 2027 takes place in San Jose, California. NVIDIA's GTC home page gives 15 to 18 March 2027, its FAQ page says 14 to 18 March, and registration had not opened on 1 November 2026. Vera Rubin NVL72 (72 Rubin GPUs, 36 Vera CPUs) is the rack you can actually get: production shipments began in August 2026 and Oracle Cloud Infrastructure announced general availability on 27 October. NVIDIA's datasheet rates it at 227 kW, 187 kW or 156 kW depending on operating mode. Rubin Ultra systems remain targeted for the second half of 2027 in the new Kyber rack, with Feynman following in 2028. In July 2026 analyst firm SemiAnalysis reported that Kyber had slipped more than 12 months to 2028; NVIDIA answered only that its roadmap was intact. Our verdict: do not hold 2027 capacity plans for the keynote. Buy or rent Blackwell Ultra and Vera Rubin on today's terms, design any new hall for liquid cooling with a phased 800 VDC path, and use GTC 2027 to get Rubin Ultra dates and rack specifications in writing.
About this article. This briefing is for infrastructure, procurement and finance leads planning AI capacity for 2027 and 2028. We checked dates and specifications against NVIDIA's pages, datasheet, press releases and earnings-call transcript on 1 November 2026, and label reported delays as reports. We have not tested the hardware. This is not legal or financial advice; the data protection notes rest on the DPDP Act 2023 and the Rules notified in November 2025. Vendors change dates and terms, and no vendor paid for inclusion. Researched with AI assistance and reviewed by our editorial team.
What NVIDIA has confirmed about GTC 2027
NVIDIA's GTC pages give 15 to 18 March, but the FAQ gives 14 to 18 March in two answers, without explanation. In 2026 its schedule opened with Sunday workshops before the Monday keynote, and 14 March 2027 is also a Sunday, so a training day is plausible but unconfirmed. Treat 15 to 18 March as the main conference and re-check when registration opens.
| Item | What NVIDIA says (1 Nov 2026) | Status |
|---|---|---|
| Dates | 15 to 18 March 2027 on the home page; 14 to 18 March in the FAQ | Day count unresolved |
| Venue and format | San Jose McEnery Convention Center and venues across San Jose; hybrid | Confirmed |
| Registration | Due to open later this autumn | Not open yet |
| Keynote | Not announced; the 2026 keynote ran on Monday 16 March at SAP Center | Monday 15 March likely, unconfirmed |
The 2026 keynote details are from NVIDIA's blog. Two earlier dates could move the story. NVIDIA reports third-quarter fiscal 2027 results on 17 November 2026, and GTC Washington, D.C. runs from 30 November to 3 December 2026, with a Jensen Huang keynote on 1 December. Neither is promised to cover Rubin Ultra, and the word Kyber does not appear in the 26 August earnings-call transcript.
The roadmap NVIDIA has published
NVIDIA's GTC 2025 recap placed Vera Rubin NVL144 systems in the second half of 2026 and Rubin Ultra systems in the second half of 2027. At GTC 2026, Jensen Huang presented Vera Rubin as seven chips and five rack-scale systems, and previewed Feynman, linked through Kyber over both copper and co-packaged optics. Jon Peddie Research's keynote report adds the years: Kyber with Vera Rubin Ultra in 2027, Feynman in 2028.
| Generation | Rack and scale | NVIDIA's stated timing | Status on 1 Nov 2026 |
|---|---|---|---|
| Vera Rubin | Vera Rubin NVL72: 72 Rubin GPUs, 36 Vera CPUs (called NVL144 in 2025) | Second half of 2026 | Shipping since August 2026 |
| Rubin Ultra | Kyber: 576 GPUs (GTC 2025 count), shown as 144 GPUs (GTC 2026) | Second half of 2027 | Official target; slip to 2028 reported; NVIDIA says roadmap intact |
| 800 VDC power | DC distribution for 1 MW racks | Full-scale production with Kyber in 2027 (stated May 2025) | Date absent from NVIDIA's current 800 VDC page |
Why the NVL numbers changed
In March 2025 NVIDIA called the 2026 rack Vera Rubin NVL144. Its product page now sells Vera Rubin NVL72 with 72 Rubin GPUs, and its platform deep dive lists two compute dies per Rubin GPU. So 144 counts dies and 72 counts packages. Kyber was presented in 2025 as holding 576 Rubin Ultra GPUs and in 2026 as 144 GPUs, which matches if each Rubin Ultra package carries four dies. Tom's Hardware described the GTC 2026 tray that way.
Two reports have since complicated this. In late June 2026 SemiAnalysis said NVIDIA had dropped the four-die design for a two-die part, which would halve compute per package. In August 2026 The Information reported that NVIDIA was testing lower-memory versions, some with 192 GB or 256 GB, some using HBM4 rather than HBM4E. NVIDIA told Tom's Hardware it is optimising compute, networking and memory on Rubin Ultra, without confirming any configuration. Ask every vendor to state packages, dies per package, HBM per package and rack power on one line. Our AMD Helios vs NVIDIA Vera Rubin rack-quote guide covers the other normalisation traps.
Vera Rubin NVL72: the rack you can buy now
Whatever happens at GTC 2027, Vera Rubin NVL72 is the system most 2027 budgets will buy or rent. NVIDIA's product page marks it as available now. These figures come from that page and the October 2026 datasheet.
| Specification | Vera Rubin NVL72 |
|---|---|
| GPUs and CPUs | 72 Rubin GPUs, 36 Vera CPUs (3,168 Olympus cores) |
| NVFP4 inference (sparse) | 3,600 PFLOPS |
| GPU memory | 20.7 TB HBM4 at 1,400 TB/s |
| Rack power | 227 kW MaxP, 187 kW MaxQ, 156 kW MaxLPS |
| Coolant inlet temperature | 45°C |
| 100 MW AI factory sizing | 40,000 Rubin GPUs, using DSX MaxLPS |
NVIDIA claims one-tenth the cost per million tokens of GB200 NVL72, up to 10x more tokens per megawatt, and mixture-of-experts training with one-quarter of the GPUs. The token figures use Kimi-K2-Thinking at 32K input and 8K output tokens, and NVIDIA marks them as subject to change. They are vendor-run. Better silicon also does not flow straight to your invoice, as our Rubin budget analysis explains.
Rack power: in the datasheet, not on the product page
The product page gives no rack power figure, but the datasheet does: 227 kW at MaxP, 187 kW at MaxQ and 156 kW at MaxLPS. NVIDIA's headline density depends on the last of those: 100 MW across 40,000 GPUs is 2.5 kW of site power per GPU, or 180 kW per 72 GPUs, leaving about 24 kW per rack for overhead at 156 kW.
MaxLPS combines dynamic power allocation, performance-per-watt tuning and 45°C cooling, and NVIDIA projects it can fit up to 40% more Rubin GPUs in the same power budget. Three points from NVIDIA's post matter to buyers.
- The Dynamic Power Software behind it was in developer preview on 24 August 2026, so confirm its status.
- The 40,000 figure is a lifecycle target. A site starting with training-heavy work runs above the MaxLPS average and can populate fewer racks at first.
- The published rack result is for an inference workload (DeepSeek-R1), where provisioned power fell from 136 kW to 101 kW per rack, giving 35% more racks in the same envelope.
Shipping and availability so far
NVIDIA said in January 2026 that Rubin was in full production, said on 31 May that the platform was ramping into full production, and told investors on 26 August that production shipments had started earlier that month. Its March launch release lists AWS, Google Cloud, Microsoft Azure, OCI, CoreWeave, Crusoe, Lambda, Nebius, Nscale and Together AI for the second half of 2026.
NVIDIA's 21 October post carries a general availability headline, but its text says Google Cloud's A5X instances reach it next month, meaning November 2026. On 27 October NVIDIA and Oracle announced general availability of Vera Rubin NVL72 on OCI, plus plans for HGX Rubin NVL8: eight Rubin GPUs with two 64-core Intel Xeon 6 CPUs and 4 TB of DDR5. That server is a smaller step for enterprises that cannot host a 72-GPU rack, though the datasheet lists the same 45°C inlet for it, so do not assume air cooling.
Rubin Ultra and Kyber: promised versus reported
What NVIDIA has said
NVIDIA's 800 VDC technical blog says that at GTC 2025 it exhibited an 800 V sidecar to power 576 Rubin Ultra GPUs in a single Kyber rack. At GTC 2026 Huang previewed Kyber with 144 GPUs in vertical compute trays, targeted for 2027, and held up its backplane on stage.
The latest public signal came at the OCP Global Summit on 12 October 2026, where NVIDIA's Ian Buck gave the keynote. NVIDIA's blog on the summit does not mention Rubin Ultra or Kyber. Two write-ups say he previewed a Vera Rubin Ultra NVL576 system spanning several racks over co-packaged optics, and one reports him saying it arrives next year. We have not checked this against a transcript, so treat it as reported.
What has been reported
In early July 2026 SemiAnalysis said the Kyber NVL144 rack was delayed by more than 12 months, into 2028, and Tom's Hardware and CNBC reported it on 6 July. The stated cause is the circuit board midplane that carries every copper NVLink connection between compute and switch trays. SemiAnalysis also said NVIDIA dropped NVL72x2, a stopgap joining two racks back to back, after large customers objected, and expects the eight-rack NVL576 configuration over co-packaged optics to slip or ship in low volume.
NVIDIA's reply to Tom's Hardware was one sentence saying its roadmap was intact. It gave no detail, and Tom's Hardware said it was unclear whether that meant the original schedule or a revised one. CNBC described NVIDIA as rejecting the report.
How to read it
Two outcomes fit the evidence. Rubin Ultra silicon could ship in 2027 in today's 72-package rack format with Kyber later, which SemiAnalysis says would cap the largest 2027 Rubin Ultra domain at what current racks deliver. Or NVIDIA shows a production-ready Kyber at GTC 2027 with firm dates. Until NVIDIA dates it, do not plan a 2027 deployment that depends on a scale-up domain larger than 72 packages.
800 VDC and liquid cooling: decisions that outlast the chip
NVIDIA's May 2025 technical post says a 1 MW rack at 54 VDC would need up to 200 kg of copper busbar. Moving row distribution from 415 VAC to 800 VDC carries 85% more power through the same conductor size and cuts copper by 45%. NVIDIA claims up to 5% better end-to-end efficiency, up to 70% lower maintenance cost and up to 30% lower total cost of ownership, and said full-scale production of 800 VDC data centres would coincide with Kyber in 2027. Its October 2025 post concedes the industry must still agree common voltage ranges, connectors and safety practices.
A Kyber slip would not necessarily hold 800 VDC back. NVIDIA's current architecture page describes a gradual move from AC, with rack-level power racks for hybrid use in existing halls. Its 12 October post describes Visionbay's K-1 platform, which phases in 800 VDC for the Vera Rubin era rather than switching on day one. A hall built in 2027 will still host racks well into the 2030s, so plan the conversion path now.
Liquid cooling is no longer optional at the top end. NVIDIA's datasheet describes Vera Rubin NVL72 as 100% liquid-cooled with a 45°C inlet. Tom's Hardware reports Kyber is liquid-cooled by default, at up to 600 kW per rack on 2025 figures. Warmer inlet water lets dry coolers do more of the work, but NVIDIA itself says chillers remain important in hot conditions and for resilience, which matters in Indian summers. Confirm the supported inlet range and the worst-hour design with your OEM.
Supply and lead times
Announcing supplier deals at GTC Berlin on 21 October 2026, Huang called demand extraordinary and capacity the constraint. The same release frames 2027 lead times.
- Substrate capacity to rise 2.5x over three years.
- Amkor interposer wafer capacity up more than 300% by the end of 2028.
- Semiconductor testing capacity doubling from 2026 to 2027; new KYEC US test facilities target high-volume production in April 2027.
NVIDIA's chief financial officer told investors on 26 August that supply should remain a bottleneck at least through the end of fiscal 2028, that memory prices were heading higher, and that executed price increases take effect in the first quarter of that year, which starts in early 2027. So expect allocation for the newest parts to stay tight in the year Rubin Ultra is meant to launch. The NVL72 page offers a sales contact rather than a list price, and the Oracle announcement carried no prices. Our AI compute capacity-planning playbook covers how to plan around this.
Questions to ask before and at GTC 2027
Ask these of OEMs, colocation providers and clouds before March, then again after the keynote.
| Topic | Question to ask | Why it matters |
|---|---|---|
| Rack power | Is the quoted kW MaxP, MaxQ or MaxLPS, and what will my workload draw? | NVIDIA prints 227, 187 and 156 kW for the same rack |
| Cooling | What inlet range, flow rate and CDU capacity are needed, and who owns the secondary loop? | Vera Rubin NVL72 is specified at 45°C inlet |
| 800 VDC path | Can this hall phase in 800 VDC without a rebuild, and when? | NVIDIA now describes gradual adoption |
| Upgrade path | Will a Vera Rubin NVL72 position accept Rubin Ultra, and in which rack format? | Kyber timing is disputed |
| Lead time | What is the committed delivery date, and what happens if it slips? | Packaging capacity keeps growing through 2028 |
| Cloud region | Is Vera Rubin available in Mumbai, Hyderabad or Delhi NCR, or only abroad? | DPDP and latency needs |
Buy now, rent, or wait: a decision table
| Your situation | Our recommendation | Why |
|---|---|---|
| Need capacity before mid-2027 | Buy or rent Blackwell Ultra or Vera Rubin NVL72 now | Rubin Ultra is not due before the second half of 2027 even on NVIDIA's schedule |
| Building a hall that opens in 2028 or later | Design for liquid cooling and a phased 800 VDC path | NVIDIA designs 800 VDC for racks of 1 MW and beyond |
| Cannot host a 72-GPU rack | Look at HGX Rubin NVL8 or cloud capacity | OCI plans an eight-GPU x86 Rubin server, still with a 45°C inlet |
| Personal data that must stay in India | Shortlist domestic GPU clouds and India regions, then check Vera Rubin dates | Indian Vera Rubin capacity is ordered, not live at scale |
| Large multi-year commitment | Add slip, substitution and price-protection clauses, and ask whether spend can move to the next generation | NVIDIA states its timing is subject to change |
Worked example: how many Rubin GPUs fit in a 10 MW hall
Take a hall with 10 MW of total facility power.
NVIDIA's reference density. NVIDIA sizes 100 MW at 40,000 Rubin GPUs, or 400 GPUs per MW. 10 MW gives 4,000 GPUs, or 55 full NVL72 racks (3,960 GPUs).
NVIDIA's rack power figures. Assume a power usage effectiveness (PUE) of 1.2, our assumption for a well-run liquid-cooled hall. IT power is 10 MW divided by 1.2, about 8,333 kW, before any network deduction. Dividing by each datasheet figure and rounding down to whole racks gives this.
| Operating point | kW per rack | Racks | GPUs (racks x 72) | Against 3,960 GPUs |
|---|---|---|---|---|
| MaxP | 227 | 36 | 2,592 | 1,368 fewer (35%) |
| MaxQ | 187 | 44 | 3,168 | 792 fewer (20%) |
| MaxLPS | 156 | 53 | 3,816 | 144 fewer (4%) |
The gap. The same hall carries up to 35% fewer GPUs if every rack is provisioned for peak. The difference is whether your operator runs NVIDIA's dynamic power management, what your workload draws, and what PUE you achieve.
The Kyber step. At NVIDIA's 1 MW design point, the same 8,333 kW of IT power feeds only eight racks. Facility power, not floor space, is the limit.
India, EU and US notes
India. Domestic Vera Rubin capacity is ordered or planned, and we found none live at scale. AM Intelligence, set up by Greenko's promoters, announced a binding order on 25 August 2026 for about 9,000 Rubin GPUs for a 30 MW AI factory in Hyderabad, with delivery in the first quarter of 2027, and on 6 October two further binding orders for 20,000 Rubin GPUs, with expansion in India and Malaysia. Yotta's CEO told Business Standard in August that it is in the process of ordering 50,000 Vera Rubin GPUs and 45,000 GB300s, with its 120 MW D4 site going live between May and August 2027. Tech Times, citing a September Moneycontrol interview, gives 40,000 of each. The counts differ, so ask for a dated allocation. For rupee rates, see our India GPU cloud pricing guide.
The DPDP Act does not impose blanket localisation: the government may restrict transfers to notified countries, and stricter sector laws still apply. The 2025 Rules let it specify data a Significant Data Fiduciary must keep in India, starting 18 months after the November 2025 notification. Check what actually binds you before paying for local capacity. Our companion piece covers IndiaAI GPU empanelment.
EU. Nebius said it would offer Vera Rubin NVL72 in the US and Europe from the second half of 2026, and NVIDIA's 21 October post says Google Cloud plans to extend A5X availability to Europe. If you may move GPU workloads between clouds, see our EU Data Act cloud switching guide.
US. NVIDIA's DSX release calls energy the biggest bottleneck for AI buildouts, citing more than 200 GW of projects waiting in US interconnection queues. Grid access, not GPU supply, may set a US go-live date.
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 roadmap news into a dated plan, from rack power and cooling assessments to hybrid connectivity, through our data centre and cloud networking services. Talk to us at /contact-us/.
Last updated: 1 November 2026.
