- Valley Recap
- Posts
- 23,700 at SC26: GPUs Aren't the Hard Part Anymore
23,700 at SC26: GPUs Aren't the Hard Part Anymore
Power, cooling, and storage are now the real constraints. Plus: Bay Area startups pulled $19B in November alone.


SC26 Chicago: The Growing Complexity of AI Infrastructure at Scale

AI infrastructure was front and center in Chicago this week as more than 23,700 attendees gathered for SC26, one of the world's largest high-performance computing conferences. Researchers, technology companies, infrastructure operators, and industry leaders came together to examine the systems powering the next generation of scientific discovery and artificial intelligence.
IgniteGTM was on the ground throughout SC26 week, connecting with the infrastructure community as the industry examined how to support increasingly demanding computational workloads.
The developments coming out of Chicago reflected a market investing heavily in computing performance while confronting the practical challenges of deploying, connecting, and operating these systems at scale.
AI Is Rewriting the Supercomputing Playbook
For decades, high-performance computing (HPC) has powered scientific simulation, climate modeling, engineering, and national research. The rapid expansion of AI is introducing new requirements into these environments, influencing everything from processor design to how data moves through computing clusters.
NVIDIA's announcement of its Rubin HPC GPU and Nemotron for Science demonstrated how these worlds are coming together. The company is developing hardware and models designed to support scientific workloads that increasingly combine simulation, analysis, and AI-assisted discovery.
AMD reinforced that direction through its growing presence in global supercomputing. Its technology powers four of the world's ten fastest systems, while the newly available Lux supercomputer at Oak Ridge National Laboratory combines EPYC processors and Instinct accelerators to support advanced scientific research.
These developments point toward computing environments expected to handle a wider range of workloads, often within the same infrastructure.
That flexibility introduces new challenges for operators. Different applications have different requirements for precision, memory, networking, storage, and scheduling. Supporting them efficiently requires increasingly coordinated infrastructure.
The pressure moves beyond the processor
The November TOP500 rankings offered an interesting view of the market's development. China's LineShine retained the top position, while only 27 new systems entered the list, marking the lowest turnover in its history.
Meanwhile, investment in commercial AI computing continues to accelerate, increasing demand for the same hardware, expertise, and infrastructure resources that traditional HPC environments depend on.
As these systems expand, their performance depends heavily on the components surrounding the accelerators.
Storage announcements at SC26 illustrated that dependency. Supermicro introduced new PCIe Gen6 all-flash server capabilities, while VAST Data announced storage quality-of-service technology tied to workload scheduling. Argonne's Aurora system also demonstrated the importance of data throughput by leading the IO500 production storage benchmark.

Networking is facing similar demands. NVIDIA highlighted a Lambda deployment connecting more than 10,000 GPUs using co-packaged optics technology intended to improve power efficiency and reliability.
Each advancement addresses a different part of the same operational challenge: keeping increasingly expensive computing resources productive.
A cluster may contain thousands of advanced accelerators, but its overall performance still depends on how quickly data reaches them, how efficiently workloads are distributed, and how reliably the supporting systems operate.
Infrastructure economics are becoming harder to separate from engineering
Power and cooling were also prominent throughout SC26.
Delta Electronics showcased liquid-cooling technologies alongside emerging 800 VDC and ±400 VDC power architectures designed to support higher-density AI systems.

These technologies reflect changing requirements for the facilities housing modern compute clusters. Greater rack density brings additional demands on electrical distribution, thermal management, and the physical infrastructure connecting systems together.
For operators, those requirements directly affect deployment timelines and economics.
Building additional capacity involves coordinating hardware availability, facility readiness, power access, cooling capabilities, and the ability to operate systems consistently once deployed.
The technical decisions made at the cluster level increasingly influence the financial performance of the entire deployment.
What SC26 Tells Infrastructure Investors About 2027
SC26 offered a useful view of how the high-performance computing and AI infrastructure markets are developing alongside one another.
Advances in processors and scientific AI are expanding the applications these systems can support. At the same time, developments in networking, storage, memory, and power infrastructure are addressing the growing demands of operating them.
For infrastructure providers, the implications are significant. The next wave of investment will require coordination across the computing stack, with greater attention to utilization, reliability, energy efficiency, and deployment economics.
The conversations and announcements coming out of Chicago will continue shaping how the industry approaches those priorities in 2027.
We'll carry many of these discussions forward at AI INFRA SUMMIT 6 on December 9 in San Francisco, where infrastructure leaders, operators, investors, and enterprise decision-makers will examine the challenges of building and scaling AI infrastructure, from power and compute economics to deployment and enterprise adoption.

Bay Area Startups Collectively Secured $19B+ in November MTD
The third week in November ended with $2.5B in startup investment, taking the 2026 YTD total over $218B. That is double 2025's full year of $209B and more than 4x 2024's $93B total. Continuing funding at the current pace would put the total just over $470B at year-end.
This week's market sell-off underscores the divergence between public market and private valuations. Nasdaq-listed companies alone lost $4 trillion, the broader market as much as $7 trillion. Meanwhile, private market valuations at all stages continue an upward trend, as does the number of private venture-backed companies, already at an all-time high.
Startups: Looking for investors, partners, possible competitors? Best deal in the valley is the LinkSV package, including Pulse of the Valley weekday newsletter and alerts – get a daily summary of just the activity that is most important to you: prospects for investors, partners and customers. Plus, Investor Analysis of the investors in companies like yours, summarized with co-investors, portfolio company execs and contact details. Founders get the investor analysis, alerts and newsletter for just $7/month ($50 value). Check it out and sign up here.
Follow LinkSV on LinkedIn to get early notice of new tech, stay on top of SV funding intelligence, and keep track of the trends, companies and investors impacting the startup ecosystem.
Early Stage:
Genesis AI closed a $500M Series A, a global robotics company building general-purpose robots with human-level intelligence and capabilities.
Noeri closed a $110M Seed, enables organizations to build, deploy, and improve specialized AI agents using their own data and operational knowledge.
Compute Desk closed a $45M Seed, builds market infrastructure to trade, clear and hedge compute.
VoicePatrol closed a $3.4M Seed, provides real-time voice chat protection, detecting and acting on severe toxicity and illegal activity for online communities.
Growth Stage:
Ramon.Space closed a $124M Series B, builds onboard computing systems that process, store, move data in orbit.
BrightAI closed a $97M Series B, empowers operators of critical infrastructure to collect and connect sensor data in real time.
Linq App closed a $90M Series B, the infrastructure to build AI agents in iMessage, covering a suite of enterprise-ready APIs for iMessage, SMS, RCS and voice channels.
Remedy Robotics closed a $62M Series B, provides endovascular intervention for stroke, heart attack, bleeding, pulmonary embolism and more.


Most AI Chips Were Built for Training.
SambaNova Built One for Inference.
The majority of AI hardware on the market started as training accelerators and got repurposed for serving. SambaNova went the other direction. The company designed its own chip, the RDU (Reconfigurable Dataflow Unit), from scratch around a dataflow architecture optimized for inference throughput and efficiency at serving time.
SambaNova is a full-stack AI inference platform company that builds custom silicon, rack-scale systems and cloud software to run production AI workloads on-prem, in neocloud environments and through its own hosted API.
Who they are & what they deliver
SambaNova was founded in 2017 by researchers from Stanford's computer architecture lab. In July 2026, the company completed the first close of a $1B Series F round led by General Atlantic, valuing the company at $11B. Earlier that year, it raised over $350M in a Series E led by Vista Equity Partners and Cambium Capital.
The same week the Series F closed, JPMorgan Chase selected SambaNova as its on-prem inference infrastructure partner, deploying SN40 and SN50 RDU systems so production AI workloads run inside the bank's own environment. At Hot Chips 2026, SambaNova presented the SN50 RDU architecture and shipped the first live disaggregated inference demo for AI agents, splitting prefill and decode across separate hardware to improve throughput and cost per token.
The product stack includes SambaRack (turnkey rack-scale inference systems), SambaStack (software for orchestrating inference workloads) and SambaCloud (hosted API access for developers who want to build on the platform without managing hardware).
Who they serve
SambaNova serves enterprises, neoclouds, sovereign AI deployments and U.S. public sector buyers that need high-throughput inference without sending data off-prem. Financial services, government and energy are core verticals. The company also partnered with TEPCO Systems to deliver energy-efficient AI infrastructure to Japan's power sector.
At our NeoCloud Summit on October 8 in San Francisco, SambaNova's Rob Naidoff moderated "The Inference Stack: From Silicon to Serving Models" alongside Syona Sarma (DigitalOcean) and Duncan Ng (Vultr). He returns to the AIS6 stage on December 9 to moderate "The Agent-Ready Data Center: Latency, Power, and the Cost to Serve," a panel on what it takes to run agentic workloads sustainably at scale.
// NEXT UP
Your Feedback Matters!
Your feedback is crucial in helping us refine our content and maintain the newsletter's value for you and your fellow readers. We welcome your suggestions on how we can improve our offering. [email protected]
Logan Lemery
Head of Content // Team Ignite
Live Dashboards on Your Production Data. No Warehouse.
Most dashboards read from a warehouse because running analytics on the production database used to mean slowing it down. TimescaleDB removes that tradeoff. Hypertables partition your busiest tables automatically as they grow. Continuous aggregates keep rollups current in the background, so a 30-day dashboard reads pre-computed results instead of rescanning raw rows. Columnar compression cuts storage up to 95%. No ETL job to babysit, no second copy of your data, no new query language. Same Postgres, same SQL, same drivers. Start on Tiger Cloud and get $1000 in credits.







