In partnership with

What AI INFRA SUMMIT 6 Is Built to Answer

Two years ago, AI infrastructure meant one thing: get GPUs. That scramble defined 2024. It does not define where we are now.

Today, the models work. The hardware exists. The capital is flowing. And yet most enterprises sit between pilot and production, stuck in a gap no amount of raw compute can close alone. The hard problems are about everything underneath and around the GPUs: running workloads reliably, making the economics work, governing autonomous systems, and turning capability into outcomes. Those are the questions AI INFRA SUMMIT 6 is built to answer.

On Wednesday, December 9, more than 1,000 executives, builders, and infrastructure operators will gather at The Midway in San Francisco. The program reflects what has actually happened in the back half of 2026 less conference agenda, more industry snapshot.

The Big Themes

Three conversations run through the day, and they are starting to sound like one.

Infrastructure. Tokens per watt has replaced peak FLOPs as the metric that matters. Agentic workloads rewrite the sizing math a single agent task can fan into 50 to 100 inference calls, pulling CPUs, memory, and storage back onto the critical path. Inference-first architecture, disaggregated prefill and decode, and heterogeneous silicon are now production requirements.

Enterprise. The question has moved past “should we do AI” to “how do we get from pilot to production, and how do we prove it was worth it.” FinOps for AI and cost per successful task are the metrics CFOs will sign off on. In regulated industries, governance, data residency, and compliance determine whether a project ships at all.

Capital and policy. Data centers are front-page news. Over $250 billion in projects have been paused or cancelled due to community pushback. Moratoriums on new construction span more than half of U.S. states. Power, policy, and permitting have become as important to AI’s trajectory as the silicon itself.

Who Will Be on Stage

The main stage opens with Victor Kuarsingh on bridging raw capacity and enterprise-grade product. AMD’s Ravi Singh follows on EPYC Venice and Helion rack architectures for agentic workloads. Cisco’s Murali Gandluru addresses security and observability across the AI stack. Lamia Youseff of Jazz Computing argues that where AI runs determines who controls the intelligence.

Panels cover neocloud economics time to power, cluster reliability, and multi-year compute structures and capital markets, with Celesta Capital, BlackRock, and SK Telecom debating whether GPUs are long-run assets or depreciating hardware. Sarah Welton of Infrastructure Masons confronts the reality that the digital infrastructure industry has tripled to 373 gigawatts in three years, and the next three will see more capacity built than the last 30.

Supermicro hosts a “time to token” panel with d-Matrix, LinkedIn, and Wafer. SambaNova moderates a session on the agent-ready data center. The breakout room features a Women in AI breakfast with Sandy Carter and Cisco’s Vimala Veerappan, a “Pilot to Production” working session with ServiceNow and Uber Freight, and a deep dive on the inference stack with Together AI and Arm.

Why It Matters Now

AIS6 sits at the intersection of every force shaping AI’s next chapter. Enterprise leaders, infrastructure providers, investors, and operators will all be in the same room confronting the constraints that determine how fast any of it can move. December 9 is the day those two sides of the same problem meet.

Below the stack, above the noise. If you are taking AI from pilot to production or building the infrastructure that gets it there this is your room.

AI INFRA SUMMIT 6 | Wednesday, December 9, 2026 | The Midway, San Francisco

Bay Area Startups Collectively Secured $17B+ in November MTD

The second week in November ended with $3.9B in fundings, fueled by billion-dollar-plus rounds for Skydio (Corporate, $1.25B) and Mirendil (Series A, $1B). M&A activity was up, with thirteen acquisitions, all but two (those two for $700M) for undisclosed amounts.

Anthropic's IPO is just days away and appears to be on track to meet or exceed expectations. Per Bloomberg, their IPO is six times oversubscribed. The company is also completing a $26B private placement, selling non-voting, class N shares to a group of tech companies, including Google, NVIDIA, Advanced Micro Devices, Broadcom, Micron Technology, Microsoft and TSMC Partners. The combined IPO and private placement proceeds could total more than $100B and give Anthropic a fully diluted valuation of $1.4T.

Startups: Looking for investors, partners, possible competitors? Best deal in the valley is the 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. Founders get the 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:

• Mirendil closed a $1B Series A, an AI research lab building self-accelerating systems that turn compute into scientific and engineering breakthroughs.

• Mount Thor closed a $45M Series A, builds cloud infrastructure for Apple hardware to advance AI research & deployment.

• Graphon AI closed a $35M Series A, an applied AI lab building Persistent Intelligence to power the next generation of AI agents and accelerate progress toward AGI.

• Agentcard closed an $8M Seed, a debit card for AI agents. Agents can use Agentcard’s API to issue cards and pay for things online autonomously.

Growth Stage:

• Skydio closed a $1.2B Corporate, the largest U.S.-based drone manufacturer and world leader in flying robots.

• Mercor closed a $750M Series D, our talent network trains frontier AI models by sharing knowledge, experience, and context that can't be captured in code alone.

• LangChain closed a $300M Series C, their open, model-agnostic harnesses give teams choice and control over how they build their agent architecture.

• Arch Systems closed a $30M Series B, builds expert data solutions that objectively observe and optimize manufacturing processes.

Most Neoclouds Rent Their Infrastructure. QumulusAI Owns the Whole Stack.

Building an AI cloud on leased data center space and financed GPUs is one approach. QumulusAI took a different one. The company owns its infrastructure end to end, from power and purpose-built data centers to the GPU clusters running customer workloads, giving it direct control over cost, deployment speed and capacity planning.

QumulusAI is a vertically integrated neocloud built to get AI compute online faster, deploying modular GPU clusters across its own U.S. data center sites and delivering capacity to inference providers and AI-native companies.

Who they are & what they deliver

QumulusAI (Nasdaq: QMLS) is led by CEO Mike Maniscalco, the former CTO of Applied Digital, who joined in September 2025 to scale the company's full-stack model. QumulusAI is an official NVIDIA Cloud Partner, deploying Blackwell B300 and B200 GPUs across data center sites in Philadelphia, Kansas City and Tennessee.

Rather than building massive centralized campuses, QumulusAI deploys smaller, modular compute clusters that can reach the market in months rather than years. The company follows what it calls an inference-first, demand-led approach: capacity is deployed against signed customer contracts rather than built on speculation.

In September 2026, QumulusAI completed its first Blackwell B300 deployment, activating an $18 million take-or-pay agreement at its Philadelphia site. The following month, it put a $71.9 million three-year Blackwell agreement into production across both its Philadelphia and Kansas City facilities. Its GPU fleet has more than tripled to over 3,000 GPUs, and cumulative signed customer contract value reached $523 million as of October 2026.

The company expanded its GPU lease financing to $124 million without issuing new equity, funding hardware acquisition through debt rather than dilution.

Who they serve

QumulusAI serves AI inference platform providers, AI-native companies and enterprises that need dedicated GPU capacity with predictable pricing and U.S.-based data residency. Its customer contracts are structured as multi-year, take-or-pay agreements that give buyers guaranteed access to compute and give QumulusAI revenue visibility.

In November 2026, QumulusAI contracted 0.5 MW of additional colocation capacity in Tennessee, expanding its geographic footprint as demand for distributed, U.S.-based AI compute continues to grow. The company has 7 MW of capacity already generating revenue, with more under deployment.

QumulusAI was in the room at our October 8 NeoCloud Summit in San Francisco, and Ryan Derroco, QumulusAI's CTO, is joining the AIS6 stage on December 9 for a panel on how neoclouds are gearing up for the next wave of demand from frontier AI labs.

Explore QumulusAI to learn how its full-stack neocloud model delivers GPU capacity from owned infrastructure across U.S. data centers.

// 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

The voice-to-text app that built its own speech model

Most dictation apps are a wrapper around someone else's speech model. Wispr Flow trained its own.

It's called Canto, and it was built for the way people actually talk to AI: thinking out loud, changing direction mid-sentence, half-whispering in an open office.

Wispr tested it on 10 hours of real dictations from more than 2,300 speakers. Canto had the lowest word error rate of every model tested, including models from Google, OpenAI, AssemblyAI and Deepgram. That means your prompts to ChatGPT, Claude and Cursor come through the way you said them.

Talk naturally in any app and Wispr Flow types clean, formatted text, 4x faster than typing.

Try Wispr Flow free