AWS re:Invent 2026: IST keynotes and the Trainium question
APJ-RAISE team · · 7 min read
AWS re:Invent 2026 runs from 30 November to 4 December 2026 in Las Vegas, and its keynotes stream free online, so you can follow it from India without a ticket. It is the 15th edition of Amazon Web Services' yearly cloud and AI conference.

The dates are easy to find. The more interesting thread is hardware. AWS has put the chief architect of its Trainium chip on a keynote stage this year, and Google also designs their own AI chips. So this guide follows one question: why is every big cloud company building its own AI chip?
AWS re:Invent 2026 keynote times in IST
About Amazon lists the keynotes in Las Vegas time. Las Vegas is on PST (UTC-8) in December and India is on IST (UTC+5:30), so we added 13 hours 30 minutes. Three of the five keynotes start at 10:00 PM IST.
- Developer keynote: Tue 1 Dec, 9:00 to 10:15 AM IST (Mon 30 Nov, 7:30 PM PST). Marc Brooker, Clare Liguori, Eric Johnson and Neha Rungta of AWS.
- CEO keynote: Tue 1 Dec, 10:00 PM to 12:00 midnight IST (Tue 1 Dec, 8:30 AM PST). AWS CEO Matt Garman, with guests Sam Altman (OpenAI), Tony Xu (DoorDash) and Barbara Karuth-Zelle (Allianz).
- Data & AI keynote: Wed 2 Dec, 10:00 to 11:30 PM IST (Wed 2 Dec, 8:30 AM PST). Swami Sivasubramanian, AWS VP of Agentic AI and Emerging Technologies, with guests from Delta Air Lines and Cox Automotive.
- Partner keynote: Thu 3 Dec, 4:30 to 5:30 AM IST (Wed 2 Dec, 3:00 PM PST). Ruba Borno and Matt Garman, with guests from Presidio and Marriott.
- Technical Innovations: Thu 3 Dec, 10:00 to 11:30 PM IST (Thu 3 Dec, 8:30 AM PST). Matt Wood (Chief AI and Technology Officer), Ron Diamant (VP and Chief Trainium Architect) and others.
Note the date jump. The Developer keynote is a Monday evening in Las Vegas but a Tuesday morning in India, and the Partner keynote lands before sunrise on Thursday. Werner Vogels, Amazon's CTO, is not on the 2026 keynote list. For chip news, the Technical Innovations slot is the one to watch.
Can you watch AWS re:Invent online for free?
Yes. The official livestream page says keynotes stream free there, with no sign-up. A free registration and sign-in adds more than 50 streamed talks and labs and lets you build a schedule. A pre-show called Road to re:Invent airs on Sunday 29 November, Las Vegas time, and Live from re:Invent covers 30 November to 3 December. The page does not say whether recordings will be posted.
AWS says over 2,200 sessions are now in its catalogue (as of 18 November 2026).
What re:Invent 2026 costs from India
The full in-person pass is $2,499. At ₹88 to the dollar, our estimate is about ₹2.2 lakh for the pass alone, before flights, a US visa and Las Vegas hotel nights. The livestream costs nothing, which is why it is the real event for most Indian students.
AWS also runs the All Builders Welcome Grant for people in their first five years in tech, plus some university students. For 2026 the official page says applications are closed. Check it again around mid-2027.
Why are companies making their own AI chips?
Most AI work, in training and in answering prompts, comes down to multiplying large grids of numbers (matrices). A general GPU is built to do many kinds of parallel work well. A custom chip can drop the parts that a cloud's own AI jobs never use. The reasons below are our reading of the public material, not an AWS statement.
The first reason is money. A cloud that designs its own chip does not pay another chip maker's margin, so it can rent compute cheaper or keep more profit. AWS describes Trainium as a purpose-built AI chip aimed at the best economics for training and inference.
Supply is the second. AI chips have been hard to buy in volume. Owning a second design means less waiting on one vendor.
Then there is control. When the same company writes the software stack, runs the data centres and designs the chip, it can tune all three together. Owning a chip does not mean dropping NVIDIA, though. The Trainium4 plan below is built to work with NVIDIA hardware.
What is AWS Trainium?
Trainium is Amazon's own AI chip. You cannot buy one. You rent Trainium servers on AWS by the hour.
The current generation is Trainium3. The product page says it gives up to 2 times the MXFP8 compute of Trainium2, with 144 GB of HBM3e memory per chip and 1.7 times the memory bandwidth. MXFP8 is an 8-bit number format that lets a chip do more maths per second at lower precision.
Software runs through the Neuron SDK, which AWS says supports PyTorch, vLLM, Hugging Face Transformers and Ray without code changes. For a student, that claim is worth testing on a small model before believing it.
Trainium vs TPU vs GPU: the shared idea
Both Trainium and Google's TPU lean on the same old idea, the systolic array. Picture a grid of tiny units. Each one multiplies two numbers, adds the result to a running total and passes values to its neighbour, in steady beats like a heartbeat. Each number gets reused many times as it travels, so the chip fetches from slow memory far less often.
AWS's Neuron docs say the Tensor Engine in NeuronCore-v2 is based on a power-optimized systolic array for matrix and convolution work. Google's 2017 TPU paper describes a matrix unit with 65,536 multiply-add cells. That number is exactly 256 times 256, so you can picture it as a square grid with 256 units on each side.
A GPU also has matrix units, but it keeps much more general hardware around them. That makes it flexible for graphics, science and new model types. The custom chips trade some of that flexibility for efficiency on the work their owners run most.
What most previews got wrong about Trainium4
As of 18 November 2026, AWS has confirmed no specific launch for re:Invent 2026. Several blog previews still write about Trainium4 pricing, release or sales outside AWS as if decided. We found those claims only in blogs, never in an AWS statement.
What is on record: at re:Invent 2025, NVIDIA said AWS is designing Trainium4 to work with NVIDIA NVLink Fusion, with no release date. The 2026 AI/ML track has sessions on Trainium3 and the Trn3 UltraServer (AIM310), Trainium for newcomers (AIM211) and Trainium on Amazon EKS (AIM311). A chief chip architect on the keynote stage hints at chip news, but treat that as a guess until the keynote airs.
Agentic AI at re:Invent 2026
Two AWS agent products appear in official re:Invent material. Bedrock AgentCore is pitched as a platform for production AI agents with any framework and any model, and sessions AIM204, AIM364 and AIM443 build on it. Kiro is an AWS tool that uses agents to plan, build, review and check software, and its site says you can start free.
New to the topic? Read what AI agents are, then multi-agent systems and how AI agents get an identity.
A simple plan for watching from India
- 1Put the IST times above in your calendar, with the correct Indian dates.
- 2Watch the CEO keynote on 1 December and the Technical Innovations keynote on 3 December, both at 10:00 PM IST.
- 3Register free if you want the streamed talks, and start with AIM211 for Trainium basics.
- 4Before you share any launch, find it in an official AWS announcement.
Google Cloud Next '27 follows on 13 to 15 April 2027 at Mandalay Bay, Las Vegas. For an academic conference instead of a vendor one, see our NeurIPS 2026 guide. APJ-RAISE students read announcements and papers together, and joining is free.
Questions people ask
- What time is the AWS re:Invent 2026 keynote in India?
- Matt Garman's CEO keynote is on Tuesday 1 December from 10:00 PM to midnight IST. The Data & AI and Technical Innovations keynotes also start at 10:00 PM IST, on 2 and 3 December.
- Can I watch AWS re:Invent 2026 for free?
- Yes. Keynotes stream free on the official livestream page with no sign-up. Free registration adds more than 50 streamed talks and labs.
- What is the difference between Trainium vs TPU?
- Trainium is AWS's AI chip and TPU is Google's. Both rent through their owner's cloud and both are built around systolic arrays for matrix maths.
- Will AWS launch Trainium4 at re:Invent 2026?
- Not confirmed. In December 2025 AWS and NVIDIA said Trainium4 is being designed, but as of 18 November 2026 AWS has announced no launch date or price.
Sources
- 1.AWS. AWS re:Invent 2026 official event page.
- 2.AWS. AWS re:Invent 2026 keynotes.
- 3.AWS. AWS re:Invent 2026 livestream.
- 4.AWS. AWS re:Invent 2026 - All Builders Welcome Grant.
- 5.About Amazon. AWS re:Invent 2026: What to know. Updated 13 November 2026.
- 6.Markus Adhiwiyogo. The attendee's guide to the AWS re:Invent 2026 AI/ML track. AWS Machine Learning Blog, 11 November 2026.
- 7.AWS. AWS Trainium product page.
- 8.AWS Neuron documentation. NeuronCore-v2 Architecture.
- 9.Jesse Clayton. AWS Integrates AI Infrastructure with NVIDIA NVLink Fusion for Trainium4 Deployment. NVIDIA Technical Blog, 2 December 2025.
- 10.Norman P. Jouppi et al. In-Datacenter Performance Analysis of a Tensor Processing Unit. arXiv, 2017.
- 11.AWS. Amazon Bedrock AgentCore.
- 12.Kiro: Ship quality software.
- 13.Google Cloud. Google Cloud Next '27 registration site.

