In partnership with

Beyond the GPU Land Grab: What We Heard at NeoCloud Summit 2026

SAN FRANCISCO, Oct. 8 – By the time emcee Brian Sparkes asked the crowd at Hotel Kabuki to claim their seats for the day, it was clear nobody had come for a victory lap. Over the next seven hours, operators, chipmakers, financiers and enterprise buyers kept circling one uncomfortable question. Getting GPUs used to be the hard part. What does it take to turn them into a business?

Ignite CEO Bill Barry set that tone in his opening. He described NeoClouds as the industry's new "center of gravity," evolving from GPU-as-a-service toward bare metal and full cloud offerings. He pointed out a mismatch in the stack: power and data centers take years to build, while agentic applications change in weeks. "What we're trying to do today is get everybody together so we can start asking the right questions, better questions."

1. Deliver what you promise

The opening panel, moderated by Supermicro's Carl Brown, started with fundamentals. Steven Hou, Head of Research at Silicon Data, argued that capital efficiency, utilization and tenant quality now decide what gets financed. He also pushed back on the hype: "This idea that everyone is deploying a swarm of AI agents. That's not happening yet." WEKA CEO Liran Zvibel said inference rewards whoever is "effective, nimble and cost effective." Tiffany Wilson, CEO of Remanie.ai, gave the line of the morning: "Commit to the capacity that you can deliver, not commit to capacity that you hope to deliver."

2. Build for change, not for today's benchmark

AMD's Nicola Tan warned that nobody can predict which models or agent architectures will win. "The biggest risk is not choosing the wrong technology," she said, "but building something that cannot adapt." Her test for an open stack was simple: if swapping one vendor forces you to replace everything, "you don't have a layer, you have a bundle."

3. Power is necessary, not sufficient

On the energy panel moderated by Dean Nelson of iMasons, Jazz Computing CEO Lamia Youseff contrasted AI's fast cycles with utility planning horizons of seven or eight years. She also raised data residency as a lasting driver of sovereign AI. Forrest Heath of Panadina made the case for building load next to generation in Colombia: "It's much easier to transmit bits than electrons." His warning to anyone selling power-first capacity: "The customers don't want the problems."

4. Training isn't over. It's a seesaw.

On a panel moderated by Sebastian Spitzer, GMI Cloud CEO Alex Yeh rejected the idea that the industry has simply moved from training to inference. He said post-training is growing alongside inference, and on the edge-versus-cloud debate he said, "It's cloud times edge... not versus." Craig Gieringer of Massed Compute said inference already accounts for more of his company's customers, and that it intends to serve both.

5. Stop pricing by the GPU-hour

Muneeb Rasool of Tensor Machines and Dean Nelson, now with Cato Digital, presented an open benchmark built around a simple idea. If you pay for the whole GPU, you have to account for the whole bill. Their tests showed GPUs on paper behaving differently in practice, and overclocking adding power draw without matching gains. Nelson argued that older hardware can still pay off when it's matched to the right model.

6. Trust is an operational deliverable

Cisco's Murali Gandluru moderated a session on security, which customers are now demanding earlier in the process. QumulusAI's Victor Ghadban described one customer whose poor isolation let a DDoS attack spill over onto its neighbors. "The responsibility is always on us," he said. Steve McKay of Massed Compute called security "a shared responsibility" that has to keep pace as workloads change.

7. The network is becoming part of compute

The afternoon brought the theme down to the hardware. Rob Naidoff of SambaNova, Syona Sarma of DigitalOcean and Duncan Ng of Vultr described agentic workloads as a "data center as a system" problem. Upscale AI's Aravind Srikumar put it bluntly: "The network is really becoming part of compute." On the disaggregation panel moderated by Drew Pletcher of Lightning AI, Cerebras' Angela Yeung and AMD's Mark Orthodoxou explained how their partnership splits inference. AMD hardware handles compute-heavy prefill and Cerebras handles bandwidth-hungry decode. They urged operators to track tokens per second per watt and output tokens per second per user together. On Infrinia SVP Rajeev Koodli's panel, Ghadban summed up what customers ask for: "Give us something that works."

8. Measure the outcome, and bring a bankable package

Vultr CMO Kevin Cochrane said enterprise AI needs CPUs, data services and regional inference as much as GPUs. "Infrastructure should be invisible," he said. On the tokenomics panel, TransUnion's Arzoo Mittal and WEKA Chief AI Officer Val Bercovici explained why cheaper tokens can still mean bigger bills. "We need to get to cost per task, cost per outcome," Bercovici said. Hosted.ai's Nilesh Shah gave inference buyers the same advice: "Look at the cost of the outcome." The closing finance panel, moderated by Cisco's Vladimirs Sazonovs, made the same point to capital markets. Matt Renner of Acasia said, "You can't finance the GPUs if you can't secure power." Neil Sheridan of Digital Alpha told operators to arrive with "a full package that's fully banked."

The bottom line

The GPU land grab isn't over, but it is no longer enough. The NeoClouds that win the next cycle will be the ones that deliver capacity on time, route work to the right silicon, protect tenants, and can show a customer, or a lender, what an outcome actually costs.

Bay Area Startups Collectively Secured $3.7B+ in October

October ended the first full week with $3.7B in startup funding, $3.2B (87%) of it in megadeals dominated by AI infra and applied AI. Cybersecurity (Armadin, $255M) and healthcare (General Medicine, $120M) occupied the remaining spots on the list.

The most remarkable funding was the $870M Series A for Typesafe AI, developers of Jev, an AI model that makes fast, structured decisions instead of generating conversational text. Remarkable because their seed round for $40M was less than one month earlier but not surprising because Jev's speed and low-cost efficiency drew immediate attention upon launch and racked up more than 25% of the Fortune 500 as customers in a very short time.

Startups: Looking for investors, partners, possible competitors? Best deal in the valley is the Pulse of the Valley weekday newsletter and alerts to keep you on top of all three plus database keyword search gives you even more prospects for investors, partners and customers. Founders get the newsletter, and alerts 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 companies, investors and executives impacting the startup ecosystem.

Early Stage:

• TypeSafe AI closed a $870M Series A, building models that provide fast, predictable, and efficient intelligence for apps requiring semantic judgment and intelligent decision-making.

• Luminal closed a $29.2M Series A, compiles AI models to give you the fastest, highest throughput inference cloud in the world.

• SafeWorld closed a $12.2M Seed, provides safety testing and simulation software that helps enterprises evaluate and deploy robots safely and responsibly.

• Minerva Humanoids closed a $10M Pre-Seed, developing Roger, a purpose-built semi-autonomous humanoid robot designed for hazardous energy operations.

Growth Stage:

• Oxide Computer Company closed a $445M Series D, building a new kind of server with the innovations of cloud hyperscale technology for easy on-premises compute infrastructure.

• Vinci closed a $250M Series B, building the intelligence infrastructure for a new era of hardware engineering.

• Voltus closed a $225M Series D, a leading DER technology platform and virtual power plant operator connecting distributed energy resources to electricity markets.

• Arena closed a $200M Series B, an open platform where everyone can access and interact with the world's leading AI models.

The Room Everyone Wanted to Be In

SF Tech Week filled San Francisco with more than 1,000 events, and community builders told us the average show rate across them was about 15%. Against that backdrop, 93% of NeoCloud Summit registrants walked through the doors of Hotel Kabuki on October 8. In a week where most RSVPs were optional, people treated this one as a commitment. They cleared their calendars, showed up early and stayed through the last session. To us, that's the clearest sign of where the AI infrastructure conversation is right now and how seriously this community is taking it.

The real energy of the day was in the conversations. Before the first session started, the expo was already full of people comparing notes. Lunch turned into dozens of small huddles, the Ignite Studio had a steady line of people waiting to sit down and talk, and the raffles drew real cheers from the crowd. Operators traded war stories with chipmakers, investors compared their theses with builders, and enterprise buyers found the people who could actually solve their problems. Because the room was carefully curated, nearly every introduction mattered. The happy hour in the Expo Lounge and the VIP dinner ran long, because nobody was in a hurry to leave.

"Best event of the year." We heard it over and over, in the hallways, at the dinner and in the messages that kept coming afterward. That feedback means a lot, and it belongs to everyone who made the day what it was. Thank you to our speakers and moderators for being candid on stage, to our sponsors and partners for believing in this room, to our crew and the Hotel Kabuki team for keeping everything running, and most of all to every attendee who showed up ready to share and to listen. We're grateful, and we'll see you at AI Infra Summit 6 on December 9.

// 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

Your SOC 2, handled end to end with AI agents.

Enterprise buyers won’t put your product near their customer data without a SOC 2 report. Twenty-one state privacy laws now hold them accountable for the vendors they share data with, so their diligence lands on you.

Sprinto gets you audit ready in 14 days, across three working sessions. AI agents connect your stack, collect the evidence auditors ask for, and close the gaps as they appear. You approve, they execute.

Sprinto also answers the security questionnaires your buyers send, using the same evidence base, so security reviews stop holding up deals.

No compliance hire. No consultant. Your auditor signs off.

Learn more and get $1,000 off.