What is checkable on 15 October
- NVIDIA GTC Berlin 2026 runs 20–22 October. Jensen Huang’s keynote is 21 October, 11:00–13:00 CEST at the Tempodrom. That is 17:00–19:00 in China. The livestream is free and does not require registration. Home: GTC Berlin. Catalog: Keynote SB61112.
- The pricing page says conference and exhibit-hall passes are sold out. Listed prices include VAT: Conference €1,785, Exhibit Hall €357, Tuesday workshop €471. The academic row is €1,053 / €215 / €471, which is not a 25% discount. Prices: pricing page.
- Vera Rubin is the 16 March platform: seven chips and five racks. NVIDIA’s NVL72 sentence versus Blackwell is one-fourth the GPUs for large MoE training, up to 10× inference throughput per watt, and one-tenth the cost per token. That is a March claim, not a Berlin parts list. Release: Vera Rubin newsroom.
- Agent sessions are already on the homepage, and those cards are all stamped 20 October, 11:00–13:00. Workshop BW61086 has its own catalog time: Tuesday 8:30–17:00, STATION-Berlin Session Room 3. Catalog: session catalog.
- As of 15 October 2026 the keynote block only says announcements will drop. It names no product. Anything below marked as “listen for” is a forecast, not a release.
The site tells you to wait for announcements, and the catalog still has no keynote script. The 16 March note already named Vera Rubin, Groq 3 LPX, and a cloud list. In Berlin, listen for whether a name turns into a ship date, and whether “one-tenth per token” gets a denominator.
The 21 October keynote: when, where, and what time in China
Three days: workshops on 20 October, conference and exhibits on 21–22 October. The keynote is only on the 21st. The FAQ places it at the Tempodrom (Möckernstraße 10, 10963 Berlin) and the conference at STATION-Berlin (Luckenwalder Str. 4–6, same postcode). Schedule: schedule at a glance. FAQ: GTC Berlin FAQ.
21 October is still CEST (UTC+2). European summer time ends on 25 October 2026, so the keynote has not fallen back to CET. China is UTC+8, six hours ahead. Pregame 9:00–11:00 CEST is 15:00–17:00 in China. The keynote, 11:00–13:00 CEST, is 17:00–19:00. Badge pickup at the Tempodrom starts at 8:00. A conference pass can sit in the room, and seating is limited. An exhibit-hall pass goes to a watch party, not a Tempodrom chair. The livestream is free. The FAQ says no registration is required, and the watch link will be on the GTC site.
| Berlin CEST | China time | What the 21 October schedule lists |
|---|---|---|
| 8:00–18:00 | 14:00–24:00 | Registration. Tempodrom badge pickup from 8:00 |
| 9:00–11:00 | 15:00–17:00 | GTC Live keynote pregame |
| 11:00–13:00 | 17:00–19:00 | Keynote, Jensen Huang, Tempodrom |
| 13:00–19:00 | 19:00–01:00 | Exhibit hall. Lunch is listed until 15:00 |
| 14:30–17:30 | 20:30–23:30 | Sessions, training labs, and certification exams |
| 17:00–19:00 | 23:00–01:00 | Exhibit hall reception |
Sessions are in English. The site says more than 100 sessions, from CUDA, accelerated computing, and agents to robotics, open models, and physical AI. On the 21st, sessions start at 14:30, so there are ninety minutes after the keynote for the hall and lunch, not a second launch. The 22nd has no second keynote: sessions run 9:00–12:00 and 13:00–17:00.
GPU: Berlin is about Vera Rubin, already published in March
The 16 March newsroom note says Vera Rubin opens the next agentic-AI step, with seven chips in full production: the Vera CPU, Rubin GPU, NVLink 6 switch, ConnectX-9 SuperNIC, BlueField-4 DPU, Spectrum-6 Ethernet switch, and the Groq 3 LPU. The five racks are Vera Rubin NVL72, a Vera CPU rack, Groq 3 LPX, BlueField-4 STX storage, and Spectrum-6 SPX Ethernet. Partners are described as starting in the second half of this year. 21 October sits inside that window.
NVL72 is written as 72 Rubin GPUs and 36 Vera CPUs on NVLink 6, plus ConnectX-9 and BlueField-4. The performance lines are NVIDIA’s: versus Blackwell, large MoE training with one-fourth the GPUs, up to 10× inference throughput per watt, and one-tenth the cost per token. Groq 3 LPX is 256 LPUs, 128 GB of on-chip SRAM, 640 TB/s of scale-up bandwidth, and up to 35× inference throughput per megawatt when paired with Rubin. Keep those lines labeled as NVIDIA claims. This article is not a benchmark.
| Already written on 16 March | Whose sentence | Berlin keynote as of 15 October |
|---|---|---|
| NVL72: one-fourth the GPUs versus Blackwell for MoE training; up to 10× inference per watt; one-tenth the cost per token | NVIDIA newsroom, not a cloud invoice | No script yet. Listen for a denominator: which Blackwell, which model, input or output tokens |
| Groq 3 LPX: up to 35× per megawatt; 256 LPUs; 128 GB on-chip SRAM; 640 TB/s; same second-half window as Vera Rubin | Same | Listen for a date or a European region, not a second announcement of the name |
| Seven chips in full production, five racks | Newsroom | The keynote block says announcements will drop and names no new product |
The GPU question is not whether a nameless extra card appears. Catalog entry SB61112 is Jensen Huang, same time and room as the homepage, and it has no chip table. If a new name shows up on the 21st, that is that day’s news. What you can prepare on the 15th is the March wording, so you can hear whether supply and the cost line actually change.
AI agents: sessions already on the agenda
The homepage already lists these talks, and every recommended card says 20 October, 11:00–13:00 CEST. Workshops that day run 8:30–17:00, so the two-hour stamp overlaps a full-day lab. The homepage may have printed one slot onto every card. Confirm Add to Calendar in the session catalog before you walk over. The table keeps only sessions whose title and speaker are on the homepage or a catalog page. It does not invent titles for names that appear only on the speaker wall.
| Session already named | Speaker | Relation to the keynote |
|---|---|---|
| Challenges of OpenClaw: AI’s Fastest-Growing Open Source Project | Carter Abdallah, NVIDIA | Agent runtime. Homepage topic: Agentic AI and Claws, beginner |
| Accelerating Humanoid Development With AI Agents | Spencer Huang, NVIDIA | Physical AI and robotics, not a data-center price list |
| Agentic AI in Production: From Open Models to Agents You Own (catalog code SB61091) | Kari Briski, NVIDIA | From open models to agents you run yourself |
| Building Open Models That Developers Can Inspect, Adapt, and Deploy | Bryan Catanzaro, NVIDIA | Open models you can inspect, change, and deploy |
| Bring Your Own Agent: Sovereign Guardrails for the Age of Autonomous Claws | Claire Knight, n8n | Bring-your-own agent and guardrails |
| Building Europe’s Autonomous Future With Open Models and Physical AI | Rev Lebaredian, NVIDIA | Europe, open models, physical AI |
| Inside Europe’s High-Performance AI Factories | Gabriella Scipione, CINECA, and others | European HPC factories, a panel |
| Deploying and Optimizing AI Inference at Scale (BW61086) | Yesika Marlen Ramirez Cardenas and Amit Kushwaha, NVIDIA | Catalog page: Tuesday 8:30–17:00, STATION-Berlin Session Room 3. Syllabus is prefill versus decode, KV cache, and NVIDIA Dynamo on Kubernetes |
The speaker wall also names Jörg Burzer of Mercedes-Benz, Robin Rombach of Black Forest Labs, Andrew McNamara of Shopify, plus LSEG, Siemens Energy, and L’Oréal. The site says the lineup is still growing, and those cards have no session title. Do not turn a company name into a shipped product or a scheduled keynote segment. If the keynote mentions agents, the list you can check beforehand is the table above, not the speaker wall.
Inference cost: one-tenth and 10× need a denominator
The three March lines closest to money are: NVL72 at one-tenth the cost per token versus Blackwell; up to 10× inference throughput per watt; and Groq 3 LPX with Rubin at up to 35× throughput per megawatt, plus “up to 10× more revenue opportunity” for trillion-parameter models. None of them is written as euros per million tokens for a named cloud, region, and model. DOCA Memos on BlueField-4 adds up to 5× inference throughput versus general-purpose storage. DSX Max-Q says 30% more AI infrastructure inside a fixed power envelope. DSX Flex says 100 GW of stranded grid power. Those are NVIDIA forward-looking lines, not next month’s invoice.
If Berlin turns those lines into purchasing language, the livestream has to supply a denominator, not repeat the multiplier. Four things to listen for: which Blackwell SKU, which model, and whether the one-tenth is input or output tokens; whether the 10× per watt is the same workload; whether 35× per megawatt leaves “second half” for a date you can order; and whether any European cloud region gets a price, rather than another reading of the AWS, Google Cloud, Azure, and Oracle names. None of the four is on the site on 15 October.
The engineering lab and the price list are different sessions. BW61086 starts from prefill versus decode, how the KV cache grows, and the memory-versus-compute trade, then deploys on Kubernetes and uses NVIDIA Dynamo for disaggregated inference and KV-aware routing. A small team’s bill is still GPU-hours or tokens in the cloud, or a whole machine, power, and idle time if you own it. Split card count from the monthly bill in the GPU article linked below. Do not multiply NVIDIA’s 10× into that table.
Cloud compute: the named clouds, and the Europe sessions already listed
The March note lists Vera Rubin cloud partners starting in the second half of this year: Amazon Web Services, Google Cloud, Microsoft Azure, Oracle Cloud Infrastructure, plus CoreWeave, Crusoe, Lambda, Nebius, Nscale, and Together AI. System makers include Dell, HPE, Lenovo, and Supermicro, among others. The note has no European region price, and it does not say Berlin will publish a new cloud rate card.
Europe is already on the catalog. You do not need the keynote to learn that there is a European track. Gabriella Scipione of CINECA is on Inside Europe’s High-Performance AI Factories. Rev Lebaredian’s talk puts open models and physical AI under a European frame. That is an HPC and physical-AI agenda, not a public-cloud GPU price list. Sovereign capacity, renting from a named cloud, and buying an NVL72 are three different purchase orders.
If you are not buying a rack, the day after the keynote you still have two prices: tokens on a cloud API, or GPU-hours. A keynote that only repeats “one-tenth” does not rewrite the list price in your region. If you size your own cards afterwards, use workload, concurrency, and context length, not the newsroom multiplier.
Passes: VAT-inclusive prices, a sellout, and who should just watch
The pricing page says Conference and Exhibit Hall passes are sold out. Payment is in euros, and the figures below include VAT. The FAQ says the event is sold out and points to a waitlist. The comparison table also has an academic row. A separate sentence says a university, government, or nonprofit email saves 25% on conference passes, and that offer cannot stack. Seventy-five percent of €1,785 is €1,338.75, not the table’s €1,053. Seventy-five percent of €357 is €267.75, not €215. The two sentences do not convert into each other. The academic workshop cell stays €471, the same as the regular rate, and the 25% sentence is about conference passes, not the workshop.
| Pass | Price including VAT | Relation to the keynote |
|---|---|---|
| Conference, two days | €1,785; academic row €1,053 | Tempodrom seat, limited. Includes sessions, labs, certification, exhibits, lunch |
| Exhibit hall, two days | €357; academic €215 | No Tempodrom seat. The FAQ says watch parties are open to this pass |
| Tuesday full-day workshop | €471; academic cell still €471 | At STATION-Berlin, not a keynote seat. BW61086 is this day |
| Watch the keynote online | €0; FAQ says no registration | Livestream. Not a room seat and not a session ticket |
The refund window is closed. The pricing page says a cancellation before 8:59 CEST on 22 September costs €25, and nothing is refunded after that. The FAQ writes the same cutoff as CET. On 22 September Berlin was still on summer time, so the abbreviations disagree, but on 15 October the window is shut either way. Substitution closed at 8:59 on 10 October. No-shows are not refunded. The card descriptor is RAINFOC*NVIDIA GTCEU26. In-person attendance alone requires age 18.
The pricing page also says live-session seats are limited and that all content is available after the event. If you did not get a pass, watch at 17:00 China time on the 21st and catch sessions afterwards. A sellout is not the same as “the content exists only in the room.” Proctored certification exams are free for attendees. That sentence is on the schedule page, and it is not a keynote product.
FAQ
Do I need to register to watch the keynote online?
No. The FAQ says virtual viewing is free and does not require registration. A Tempodrom seat needs a conference pass, seating is limited, and badge pickup starts at 8:00. An exhibit-hall pass is for a watch party.
Will 21 October announce a new GPU?
As of 15 October the catalog has no keynote script. The seven Vera Rubin chips and five racks are the 16 March newsroom note. In Berlin, listen for a ship date and for the denominator on the cost line. If a new name appears that day, the livestream is the source. Do not rewrite March as a Berlin debut.
The homepage puts every agent talk at 11:00 on 20 October. Is that the real slot?
Those recommended cards do all say that two-hour block, and it overlaps Tuesday’s full-day workshops. BW61086’s own catalog page says 8:30–17:00 in Session Room 3. For the other talks, use the session catalog before you move rooms.
Is “save 25%” the same price as academic €1,053?
Both sentences are on the pricing page, and they are not the same number. Twenty-five percent off €1,785 is €1,338.75. The academic conference cell is €1,053 and the exhibit cell is €215. Use the checkout page before you pay.
Can I still refund or transfer a pass?
On the pricing page, nothing is refunded after 8:59 on 22 September. Substitution closed at 8:59 on 10 October. The FAQ and the pricing page use CET and CEST for that September cutoff.
Four things to check during the livestream
These four lines are a forecast checklist. They are not announcements that have already happened.
- At 17:00 China time, open nvidia.com/en-eu/gtc/ and see whether the free watch link has become a player.
- Listen for a denominator on three multipliers: one-tenth per token, 10× per watt, 35× per megawatt. Without a denominator, it is still the March note.
- Listen for whether “second half of this year” becomes a date, a SKU, or a European cloud region. A repeated name is not a new rate card.
- Treat agent times as catalog times. The homepage’s single 11:00–13:00 slot is not a reschedule notice.
What to watch next
On 15 October the checkable items are the clock, the VAT-inclusive prices, the sellout, the 16 March Vera Rubin sentences, and the agent and inference sessions already listed. The keynote highlights are not a script yet. “Announcements will drop” says the stage will announce something. It does not say which chip.
Hear GPU, agents, inference cost, and cloud compute as separate tracks. For GPUs, start from March’s seven chips. For agents, start from the catalog, not the speaker wall. For cost, wait for a denominator. For cloud, wait for a region and a date. The CINECA panel is already listed. It does not appear for the first time if the keynote mentions Europe.
After the keynote, an agent still needs a machine that stays on
This keynote is about data-center GPUs and NVIDIA’s per-token wording. A cloud Mac mini does not replace Vera Rubin, and it does not compute that one-tenth. It is a different order: a coding agent, a repo, and a long session need a Unix machine that does not sleep. Apple silicon idles near 4W. macOS already has a shell, Homebrew and SSH work out of the box, and the machine can hold a build or an agent overnight without buying a workstation for one launch. kvmboot rents a cloud Mac mini by the day. Power and the machine sit inside the rental.