NVIDIA GTC Berlin 2026 runs 20–22 October. The keynote is 21 October, 11:00–13:00 Central European Summer Time (CEST) at the Tempodrom in Berlin, delivered by NVIDIA CEO Jensen Huang. The conference and the exhibit hall are at STATION-Berlin. On-site content is in English. Times, prices, and sessions come from the schedule, FAQ, pricing page, and session catalog as opened on 15 October 2026. When you add a session, use the room printed in the catalog that day.
The badge is an October invoice. Inference sits inside the API token price, or on a GPU you rent yourself. An agent that has to stay up overnight, or a step that has to run on macOS, is the line that adds a machine which does not sleep. On the day this was written, the plan page listed Mac mini M4 · 16GB / 256GB from $101.7 a month and 24GB / 512GB from $200.3, both with 1Gbps and a dedicated IPv4. Checkout follows the plan page.
What each of the three days is for
The FAQ splits the program into four lines: training workshops on the 20th, the keynote on the 21st, the conference on the 21st and 22nd, and exhibits on the 21st and 22nd. The schedule page has the clock times. Talks and panels are on the conference pass. An exhibits-only pass does not include them.
| Tuesday 20 October | Wednesday 21 October | Thursday 22 October | |
|---|---|---|---|
| Place | STATION-Berlin | Keynote at the Tempodrom; the rest at STATION-Berlin | STATION-Berlin |
| Clock (CEST) | Workshop registration 7:00–8:30; workshops 8:30–17:00; lunch 12:00–13:00 | Registration 8:00–18:00; pregame 9:00–11:00; keynote 11:00–13:00; exhibits 13:00–19:00; sessions and labs 14:30–17:30; reception 17:00–19:00 | Registration 8:00–13:00; exhibits 8:30–18:00; morning sessions 9:00–12:00; afternoon sessions 13:00–17:00; labs until 18:00 |
| The session that touches the bill | Inference workshop BW61086 | Keynote; SB61091 in the afternoon | Humanoids and agents, SB61076 at 13:00 |
Both addresses are in the FAQ. Conference: STATION-Berlin, Luckenwalder Str. 4–6, 10963 Berlin. Keynote: Tempodrom, Möckernstraße 10, 10963 Berlin. The GTC homepage says more than 100 sessions, from CUDA, accelerated computing, and agentic AI to robotics, open models, and physical AI. The schedule describes sessions with the same set of words: physical AI, AI factories, agentic AI, and inference.
20 October: the inference workshop
The full-day session that speaks directly to the inference bill is Deploying and Optimizing AI Inference at Scale [BW61086]. The catalog lists 8:30–17:00 CEST on 20 October at STATION-Berlin, Session Room 3. The speakers are NVIDIA solutions architects Yesika Marlen Ramirez Cardenas and Amit Kushwaha. The takeaways are three lines: reason about LLM inference from first principles, including prefill versus decode, how the KV cache grows with context, and the memory–compute trade-off; deploy and benchmark on Kubernetes; use NVIDIA Dynamo for disaggregated inference with KV-aware routing and scheduling.
The infrastructure agenda prints the same session at Hall 4 L. Both pages put it at 8:30–17:00 CEST on Tuesday. Use the room in the catalog when you add it. The training page says in-person full-day workshops are sold out, and public virtual workshops remain, across time zones. On the pricing page the Tuesday workshop is €471 including VAT, and the card is a waitlist.
21 October: keynote and agents
Jensen Huang’s live keynote is at the Tempodrom, 11:00–13:00 CEST. The GTC Live Keynote Pregame is 9:00–11:00. The FAQ says badge pickup at the Tempodrom opens at 8:00. Seating is limited and requires a conference pass. An exhibits-only pass goes to a watch party, not into the keynote hall. The same FAQ says the livestream is free and does not require registration.
That afternoon, watch Agentic AI in Production: From Open Models to Agents You Own [SB61091]. The catalog says 14:30–15:10 CEST at STATION-Berlin, Session Room 4. Kari Briski (VP, AI Software Product Management) and Justin Boitano (VP of Enterprise and Edge Computing). The takeaways: new open models, tools, skills, and secure runtimes in the NVIDIA Agent Toolkit; how to build, deploy, and optimize specialized agents; examples of making them faster, safer, and more efficient. The open-models page lists the same session as 8:30–9:10 a.m. EDT, which is 14:30–15:10 CEST, matching the catalog clock, and prints the place as Hall 2. Use the room in the catalog.
The exhibit hall reception is 17:00–19:00. The schedule labels that block’s proud sponsor as Google Cloud. That is a sponsorship credit for the reception, not a cloud invoice.
22 October: sessions
The 22nd has no second keynote. Sessions run 9:00–12:00 and 13:00–17:00, and labs run until 18:00. The catalog places Accelerating Humanoid Development With AI Agents [SB61076] at 13:00–13:40 CEST, Session Room 5. That session puts agents into the humanoid development loop. The GPU on the simulation and robotics side stays on the GPU line. It does not move onto the Cloud Mac monthly rent.
On the pricing page, exhibits-only for two days is €357 including VAT. The list has the pregame and watch party, Expo Theater, Connect With the Experts, certifications, exhibits, receptions, and lunch. Talks and the Tuesday workshop sit on the conference pass and the workshop column.
GPUs and inference capacity
The schedule does not publish an hourly price for a GPU. BW61086 is about the fact that one hourly rate still produces two different GPU-hour totals for prefill and for decode. Before you host the weights yourself, measure the two stages apart.
Prefill, decode, and the KV cache
One request splits in two. Prefill consumes the prompt and is compute-heavy. Decode writes one token at a time, and the KV cache occupies GPU memory; a longer context hits memory first. Disaggregated inference puts the two stages on different GPUs, then uses KV-aware routing to choose where the request goes. Dynamo is the scheduler this workshop names.
- API calls only. The GPU is already inside the token price. Use the 12 October agent bill or the 14 October Argon prices for the model line. When the weights are not on your own machine, leave the extra GPU off that invoice.
- You host the weights. Replace the model line with GPU rental. See whether prefill is the queue, or the KV cache is what fills memory. The hourly rate can be the same while the GPU hours differ by a lot. The workshop’s Kubernetes benchmark is that measurement. GTC does not publish the hourly rate, and this article does not invent one.
- Mac mini M4 is a different line. The plan page is a dedicated bare-metal Mac, 16GB or 24GB of unified memory. It holds the agent process, and models that fit in that memory. A gigawatt AI factory stays on the GPU line. For whether 16GB or 24GB is enough for on-box inference, see Mac mini M4 for AI development. For scheduling a self-hosted model, see DeepSeek inference performance.
Four things an AI agent bills separately
The title of SB61091 is agents you own. The four things in the takeaways get four looks at the invoice: open models, tools, skills, and a secure runtime. The model is tokens or a GPU you host. Tools are search, a sandbox, or a command that has to run on macOS. If skills and the runtime are only software, and the invoice has no such line, the monthly total has no such line.
The 12 October article already ran one base volume: 880 sessions, four rounds each, 52.8M input, 2.64M output. GPT-6 Luna’s model line is $6.60. With search and a 5-minute 1GB sandbox, that sheet is $22. When the command has to run on macOS, add the 16GB Cloud Mac and the total is $123.70. This article leaves the GTC badge outside those three lines and does not recompute the tokens.
SB61076 puts agents into humanoid development. That machine is the GPU on the simulation and robotics side. A Cloud Mac shows up on the step that cannot leave macOS, such as xcodebuild, a simulator, or signing.
Cloud server cost
The badge
On 15 October 2026 the pricing page said payment and the receipt are in euros, the prices below include VAT, and VAT is called out again at checkout. The top of the page says conference and exhibit hall passes are sold out. All three purchase cards are a waitlist. The FAQ says the event is sold out and points to the waitlist. This article does not convert euros into dollars.
| Conference, two days | Exhibits only, two days | Full-day workshop, Tuesday | |
|---|---|---|---|
| Regular rate (labeled through 19 October) | €1,785 | €357 | €471 |
| Academic row | €1,053 | €215 | €471 |
| A seat at the live keynote | Yes, limited | Watch party | No |
| Talks and panels | Yes | No | This column is the workshop itself |
| Certifications | On the list | On the list | The workshop list is the eight-hour workshop and lunch |
The discount section says a university, government, or nonprofit email saves 25% on conference passes, credentials are checked at badge pickup, and the offer cannot be combined. The comparison table’s Academic row is €1,053. Seventy-five percent of €1,785 is €1,338.75, a different number from €1,053. Pay the figure the pricing page shows at checkout. The training page says on-site proctored exams are free, with a stated value of €115–€425, including the NVIDIA Certified Associate: AI Model and Application Development exam launching at GTC Berlin. When the pass list already includes certifications, that €115–€425 stays inside €1,785 or €357. Add it as its own line only when you pay for the exam without a pass.
Cancellation: the pricing page says a cancellation before 8:59 a.m. CEST on Tuesday 22 September 2026 costs a €25 fee, and there is no refund after that. The FAQ writes the same clock time as CET. Substitution closed at 8:59 on 10 October. Today is 15 October, so both deadlines have passed. No-shows are not refunded.
Three invoice lines
The monthly total follows 12 October: model tokens, plus tools that actually ran, plus the machine that stays on. If you are at the laptop and it sleeps when the lid closes, the machine line is $0.
| Line | Base used here | When this line is $0 |
|---|---|---|
| Model / API | Luna $6.60. Argon uses the two date columns from 14 October | No calls this month |
| Tools | After search and a 5-minute 1GB sandbox, the Luna sheet is $22 | No tool calls |
| Always-on machine | 16GB $101.7, or 24GB $200.3 | The lid closes the work, or every tool is an API and a sleepless machine is already there |
| GTC badge | €1,785 / €357 / €471, VAT included | You watch the free livestream and buy no pass |
$22 + $101.7 = $123.70. While that Mac is awake, Luna’s output is still priced at the 12 October rates. Rent and tokens stay on two lines. One command is one machine: if it runs in the vendor sandbox, bill the sandbox minutes and keep the machine line at $0; if it is xcodebuild, a simulator, or signing, bill the Cloud Mac and write the sandbox line as 0. Both lines stay only when two machines do different work.
When you host the weights, GPU rental replaces the model line. It does not sit beside the $6.60. Add-ons on the plan page change only the machine line: +1TB storage $14.3 / month, +2TB $28.6 / month, parallel link $8.6 / month. Regions are Singapore, Japan, South Korea, Hong Kong, US East, and US West. The token price does not change with the region.
The two-day conference pass is one charge. November’s agent bill has the model, the tools, and the always-on machine. After the pass sells out, the free item is the keynote livestream, not GPU hours.
What is left after the event
- The keynote livestream. The FAQ says it is free and needs no registration. A seat at the Tempodrom needs a conference pass, and seating is limited.
- Content after the event. The pricing page says live-session access is limited, and that all content is available at your convenience after the event. That sentence does not say a pass is unnecessary, and it does not give a date. For the evening of the 22nd, the free item the FAQ names is that livestream.
- Inference. If you stay on an API, keep the 12 October and 14 October sheets. If you host the weights, run the Kubernetes benchmark BW61086 describes, then fill in GPU hours. Ask your GPU provider for the hourly rate.
- The always-on machine. If the agent runs overnight, or a tool has to be macOS, the machine line is the plan-page rent. If the process only sends HTTPS, every tool is an API, and a sleepless machine is already there, the machine line stays $0.
FAQ
Does the keynote need a ticket?
The livestream is free and does not require registration. A seat inside the Tempodrom requires a conference pass, seating is limited, and badge pickup at the venue opens at 8:00. An exhibits-only pass goes to a watch party.
Can you still buy a pass on 15 October?
The pricing page says conference and exhibit hall passes are sold out, and all three purchase cards are a waitlist. The training page says in-person full-day workshops are sold out, and public virtual workshops remain. Whether the waitlist releases a pass depends on the waitlist page that day.
Does €1,785 include VAT?
Yes. The pricing page says every price includes VAT, and VAT is listed again at checkout. The Academic row is €1,053, and the discount section separately says conference passes save 25%. Seventy-five percent of €1,785 is €1,338.75. Pay the figure on the page at checkout.
Does the badge include GPU hours?
No. €1,785, €357, and €471 are the badge. GPU hours are the rental when you host the weights, or they are already inside the API token price. Cloud Mac rent is the always-on machine line.
Can a Cloud Mac replace that GPU?
It does not replace the data-center GPU in BW61086. A 16GB or 24GB Mac mini M4 runs the agent process and models that fit in that unified memory. When the weights need a data-center GPU, the charge is GPU rental or the token price. The macOS toolchain is what adds $101.7 or $200.3 to the month.
Match the schedule, then the three invoices
On the 20th, the inference workshop is prefill, decode, and Dynamo. On the 21st, the keynote is at 11:00 and SB61091 is at 14:30. On the 22nd, SB61076 is at 13:00. €1,785 stays in October. $101.7 is added only on the step that cannot leave macOS. The token price and the monthly rent stay on two lines.
Leave the overnight agent on a Mac that does not sleep
The laptop is where you match one GTC session to these three lines. The overnight loop, and the step that has to run on macOS, go on a Cloud Mac that does not sleep. The badge, the tokens, and the rent each keep their own line.