
Sign up to save your podcasts
Or


Cerebras IPO is the only thing to talk about this week. 🔥
IPO prices at $185/share. Pops nearly 70% right after. The first wafer-scale chip company to make it public — after a 40-year curse killed every prior attempt.
A water-cooler-style convo on what Cerebras actually builds, why a 23 kW wafer is a power and cooling nightmare, why 44 GB of SRAM is both the magic and the wall for LLM inference, and the cursed Trilogy Systems saga that Gene Amdahl tried — and failed — to pull off in 1983.
Why does Cerebras leave the whole wafer intact instead of dicing it? How do they route around defects to harvest ~900K working cores out of ~1M? Why is power delivery vertical, and why does the wafer literally expand a tenth of a millimeter when it heats up? What does the OpenAI deal actually buy — wafers, or tokens? And why does that distinction matter?
Chapters:
0:00 Cold open: 23 kW per wafer
0:15 Cerebras IPO day at $185
2:39 What's a wafer-scale engine
10:30 Power, cooling, and thermal expansion
18:12 The 44 GB wall
26:35 The Trilogy Systems curse
32:11 Supercomputing → training → inference
39:36 The OpenAI deal and the Wild West
Relevant reading:
Vik's Substack post on the Cerebras IPO and OpenAI deal: https://www.viksnewsletter.com/
Follow Chipstrat:
Newsletter: https://www.chipstrat.com
X: https://x.com/austinsemis
Follow Vik:
Newsletter: https://www.viksnewsletter.com/
X: https://x.com/vikramskr
Follow Semi Doped:
Get more of Austin and Vik daily, free!
Sign up: https://www.semidoped.com/
Gimlet Labs runs an inference cloud built on heterogeneous silicon. Their software traces a PyTorch workload, segments it into its component parts, and schedules each piece onto the best-suited hardware — connecting chips from different vendors on a single high-speed fabric.
In this interview, Gimlet co-founder Natalie Serrino and former Intel executive Beltir walk through the architecture (graph trace, optimal split points, lowering each segment to TensorRT on NVIDIA and equivalents elsewhere), the three customer segments they sell into (frontier labs, sovereign clouds, AI natives), and a concrete demo: on GPT-OSS 120B at 8K input / 1K output, running the speculative decoder on a d-Matrix Corsair card while NVIDIA B200s handle the verifier shifts the throughput-vs-interactivity Pareto frontier roughly 4× over GPU-only speculative decode.
The most surprising takeaway: most Neoclouds gave significant equity to a single silicon vendor in exchange for capacity. Hardware amortization is around 70% of their annual costs, and the equity terms prevent them from diversifying their silicon. So the only software innovation they can ship is disaggregation on top of one vendor's stack — never across vendors. Gimlet's two-track model (deploying orchestration software inside customer data centers, plus running their own Neocloud built on mixed silicon) is the answer to that constraint.
Read the full transcript on Chipstrat.
Chapters:
0:00 Intro and the chips no one's connected before
0:33 Inference cloud for agents
1:02 From Intel to Gimlet
2:14 The case for heterogeneous inference
4:03 Disaggregating inference by resource profile
6:24 Tracing PyTorch into a schedulable graph
8:08 Connecting chips never connected before
10:52 CPUs as the agentic workhorse
12:01 Tool calls in the same data center as the LLM
13:21 Latency vs throughput on a shared fabric
14:57 Three customer buckets
15:54 Sovereigns: make an API call, not a porting project
19:37 "Cracked software is the platform"
22:24 Why merchant silicon vendors need partners
25:18 Hyperscalers outsourcing CapEx, not just kernels
28:49 AI natives: latency budgets, not just price
32:06 The d-Matrix partnership
33:31 The Pareto frontier chart
35:56 Speculative decode on Corsair: 4× shift
37:27 4× faster, or 3× more customers?
41:22 Why most Neoclouds can't follow this model
42:34 Gimlet's two-track business model
44:30 CoreWeave vs Together vs Gimlet
45:15 Series A and hiring
Relevant reading:
The Information on Gimlet helping OpenAI optimize for Cerebras: https://www.theinformation.com/newsletters/ai-agenda/startup-helping-openai-optimize-ai-cerebras-chips
Sachin Katti and Zain Asgar coauthored research at Stanford: https://arxiv.org/abs/2507.19635
Follow Chipstrat:
Newsletter: https://www.chipstrat.com
X: https://x.com/chipstrat
What's common to optics and power that ruins everything in the era of AI?
Resistance.
The same physics that drove interconnects to optics is now driving low-voltage power delivery up to 800V. Austin Lyons (Chipstrat) and Vik Sekar (Vik's Newsletter) unpack it using the Kyber rack as an example.
At 600kW and 48V, you're pushing 12,500 amps through a single rack. Power loss scales with I². The math doesn't work. The fix is 800V — and the parts come straight from the EV traction inverter ecosystem (SiC, GaN, IGBTs).
We cover the full grid-to-GPU power conversion chain (substation, utility room, PSU, intermediate bus converter, VRM), why vertical power delivery is the CPO equivalent for power, and why the power industry is a much wider open problem than optics or HBM. Plus the new topology fight: 800V → 48V (reuse the existing 48V infrastructure) vs 800V → 6V (skip 48V entirely, like TI and Navitas are pushing).
We also touch Coherent's six-inch indium phosphide ramp at Järfälla, Sweden, and why margins are the real read-through next quarter.
Relevant reading:
Vik's Substack post on power: https://www.viksnewsletter.com/p/power-delivery-as-the-next-physics-wall
Google TPU 8i / 8t blog (Boardfly deep dive): https://cloud.google.com/blog/products/compute/tpu-8t-and-tpu-8i-technical-deep-dive
Get more of Austin and Vik daily, free! Sign up here: https://www.semidoped.com/
Follow Chipstrat:
Newsletter: https://www.chipstrat.com
X: https://x.com/austinsemis
Follow Vik:
Newsletter: https://www.viksnewsletter.com/
X: https://x.com/vikramskr
Chapters
(00:00) Intro
(01:41) Memory tax: inflation, not innovation
(03:46) Boardfly: 16 hops to 7
(05:12) Coherent's six-inch indium phosphide ramp
(12:15) Power is the next physics wall
(15:08) Why 48V breaks at 600kW: 12,500 amps
(23:05) 800V and vertical power delivery: CPO for power
(30:34) Grid to GPU: every stage is a different supply chain
(39:20) 800V → 48V or skip straight to 6V?
The hyperscaler memory tax quarter.
More CapEx? Pssh. We knew flops needed scaling.
But $25B at Microsoft alone just to pay higher component prices?
A memory tax. That's the news.
NAND? Sold out. HBM? Sold out.
What we cover:
New Semi Doped with @vikramskr and @austinsemis.
Check out our Substacks
- https://www.viksnewsletter.com/
- https://www.chipstrat.com/
Chapters:
0:00 Intro and Vik goes full-time
5:15 Earnings week: the memory tax
7:26 Samsung HBM4 and the Gbps race
14:42 Is the memory tax worth it?
17:37 SanDisk and the SunDisk origin
23:22 78% gross margins and 5-year supply lock-ins
29:29 DeepSeek v4 and SSD-centric inference
38:49 Hyperscaler CapEx and the cloud pull
42:49 AI accelerators: TPU, Trainium, MTIA
Google's Cloud Next 2026 keynote? Fire. 🔥
The TPU is now two chips instead of one — 8t for training, 8i for inference — but more interestingly, it's two scale-up networking topologies too.
Austin Lyons (Chipstrat) and Vik Sekar (Vik's Newsletter) walk through what actually changed, one day after the announcement. OCS? Yes. AECs? Yep. Copper? Yep. Optics? Yep.
We cover Virgo (Google's 47 petabit/second scale-out fabric, built entirely on OCS), Boardfly (the new scale-up topology for MoE inference that cuts hop count from 16 to 7), and the 3D torus Google still uses for training.
Why is optical circuit switching the substrate of Google's data center? Why do active electrical cables still carry scale-up traffic inside racks? Why did Google split the CPU layer too, with custom ARM Axion head nodes to keep the TPUs fed?
Along the way we trace the Dragonfly topology lineage to a 2008 paper by John Kim, Bill Dally, Steve Scott, and Dennis Abts. Abts went on to build Groq's rack-scale interconnect before landing at Nvidia.
Chapters:
0:00 Intro
0:21 Two TPUs for two workloads
2:31 HBM, SRAM, and Axion CPUs
7:22 Why networking is the new bottleneck
17:14 Virgo: rebuilding scale-out on optics
25:24 3D torus Rubik's Cube scale-up for training
34:50 Boardfly: scale-up for MoE inference
42:07 Workload-specific everything
Follow Chipstrat:
Newsletter: https://www.chipstrat.com
X: https://x.com/austinsemis
Follow Vik:
Newsletter: https://www.viksnewsletter.com/
X: https://x.com/vikramskr
Matt Steiner, VP of Monetization Infrastructure, Ranking & AI Foundations at Meta, walks through how Meta's ad system actually works, and why the infrastructure behind it differs from what you'd build for LLMs.
We cover Andromeda (retrieval on a custom NVIDIA Grace Hopper SKU Meta co-designed), Lattice (consolidating N ranking models into one), GEM (Meta's Generative Ads Recommendation foundation model), and the adaptive ranking model, a roughly one-trillion-parameter recommender served at sub-second latency.
We get into why recommender workloads aren't embarrassingly parallel like LLMs (the "personalization blob"), what that means for Meta's MTIA custom silicon roadmap, and how LLM-written kernels (KernelEvolve) flipped the economics of running a heterogeneous hardware fleet. Demand for software engineering has actually gone up as the price has come down. Meta now wants ~100x more optimized kernels per chip.
Read the full transcript at https://www.chipstrat.com/p/an-interview-with-meta-vp-matt-steiner
Chapters:
0:00 Intro and scale
0:39 How Meta's ad system works
2:00 Meta Andromeda and the custom NVIDIA SKU
3:30 Lattice: consolidating ranking models
5:00 GEM, Meta's ads foundation model
6:30 Adaptive ranking for power users
8:17 The scale: 3B DAUs at sub-second latency
9:40 Why longer interaction histories matter
10:45 The anniversary gift analogy
12:57 A decade of compute evolution
15:21 Meta's infra as a CP-SAT problem
16:07 Co-designing Grace Hopper with NVIDIA
17:47 Matching compute shape to workload
18:26 Influencing hardware and software roadmaps
20:23 MTIA: why ads aren't LLMs
22:07 The personalization blob and I/O ratios
26:38 One trillion parameters at sub-second latency
28:26 Heterogeneous hardware trade-offs
29:30 KernelEvolve: LLMs writing custom kernels
33:30 GenAI and recommender systems cross-pollination
35:21 The 2-year infrastructure outlook
37:00 Why demand for software engineering is rising
38:53 How Matt stays on top of it all
Relevant reading:
KernelEvolve (Meta Engineering): https://engineering.fb.com/2026/04/02/developer-tools/kernelevolve-how-metas-ranking-engineer-agent-optimizes-ai-infrastructure/
Follow Chipstrat:
Newsletter: https://www.chipstrat.com
X: https://x.com/chipstrat
Austin and Vik discuss Credo's acquisition of Dust Photonics, XPO as the new standard for scale-out (maybe instead of CPO?) and some thoughts about Nuvacore entering the CPU scene for agentic AI.
Gavin Baker's tweet: https://x.com/GavinSBaker/status/2044410644301046031?s=20
Vik's Substack: https://www.viksnewsletter.com
Austin's Substack: https://www.chipstrat.com
Chapters
00:00 Introduction to the Semiconductor Landscape
02:49 The Rise of Nuvacore and CPU Innovations
05:27 The Demand for CPUs in the AI Era
07:59 Photonics: The Next Frontier in Semiconductors
10:26 Credo's Acquisition of Dust Photonics
13:12 Vertical Integration in Semiconductor Companies
15:15 The Future of Copper and Optical Technologies
20:28 The Evolution of AI Training Models
25:28 Innovations in Optical Interconnects
31:10 The Future of Data Center Connectivity
36:56 Strategic Implications in the Optical Ecosystem
In this episode, Austin and Vik discuss if Intel is finally back with CPU partnerships with Google, and heterogeneous inference with SambaNova, while market cap soars above $300B. Vik tries to get his OpenClaw instance to dream every night.
Chapters
00:00 Anthropic's New Direction: Chip Development
02:30 Navigating Subscription Changes and Token Costs
05:25 Exploring Alternative AI Models
08:10 The Economics of AI: Rent vs. Buy
10:56 Intel's Resurgence and Market Dynamics
15:23 Intel's Strategic Partnerships and Market Positioning
19:37 The Role of IPUs in Modern Computing
25:08 Coexistence of x86 and ARM Architectures
29:55 Innovations in Chip Architecture and Future Prospects
Reiner Pope is the co-founder and CEO of MatX, the startup building chips designed from first principles for LLMs. Before MatX, Reiner was on the Google Brain team training LLMs, and his co-founder Mike Gunter was on the TPU team. They left Google one week before ChatGPT was released.
A counterintuitive throughput insight from the conversation:
“Low latency means small batch sizes. That is just Little’s law. Memory occupancy in HBM is proportional to batch size. So you can actually fit longer contexts than you could if the latency were larger. Low latency is not just a usability win, it improves throughput.”
We get into:
• The hybrid SRAM + HBM bet, and why pipeline parallelism finally works
• Overcoming the CUDA moat
• Why frontier labs are willing to bet on an AI ASIC startup
• Memory-bandwidth-efficient attention, numerics, and what MatX publishes (and what it does not)
• Why 95% of model-side news is noise for chip design
• Why sparse MoE drives MatX to “the most interconnect of any announced product”
• How MatX uses AI for its own chip design
• The biggest challenges ahead
Chapters:
00:00 “We left Google one week before ChatGPT”
00:24 Intro: who is MatX
01:17 Origin story: leaving Google for LLM chips
02:21 GPT-3 and the “too expensive” problem
04:25 Why buy hardware that is not a GPU
05:52 Overcoming the CUDA moat
08:46 Early investors
09:35 The name MatX
09:59 The chip: matrix multiply + hybrid SRAM/HBM
12:11 Why pipeline parallelism finally works
14:22 Reading papers and Google going dark
15:20 Research agenda: attention and numerics
17:06 Five specs and meeting customers where they are
19:24 Why frontier labs are the natural first customer
20:32 Workloads: training, prefill, decode
22:18 Little’s law and the throughput case for low latency
24:29 Interconnect and MoE topology
26:35 Inside the team: 100 people, full stack
28:32 Agentic AI: 95% noise for hardware
30:35 KV cache sizing in an agentic world
32:11 How MatX uses AI for chip design (Verilog + BlueSpec)
34:23 Go to market: proving credibility under NDA
35:12 Porting effort for frontier labs
36:34 Biggest skepticism: manufacturing at gigawatt scale
37:32 Hiring plug
Austin Lyons @ Chipstrat: https://www.chipstrat.com
Vik Sekar @ Vik's Newsletter: https://www.viksnewsletter.com/
Intel Foundry just partnered with Elon Musk’s Terafab. What is Terafab anyway, why vertically integrated fabs make sense but the economics don’t (yet!), and what Intel is doing here (hint: no idea).
Then: OpenAI acquires TBPN for an estimated $100-300M. Not sure why, but the more interesting thing is the value of niche audiences when five companies control a trillion dollars in AI capex.
And finally, Citrini Research sent an analyst to the Strait of Hormuz with a Pelican case full of spy gear, $15K cash, and Cuban cigars. The most unhinged research trip in Substack history.
Austin Lyons — Chipstrat (https://chipstrat.com) Vik Sekar — Vik's Newsletter (https://www.viksnewsletter.com)
Subscribe for weekly episodes on semiconductors, AI, infrastructure, and the business of chips.
From the publisher's feed
The business and technology of semiconductors. Alpha for engineers and investors alike.