Attending Day 2 of SF Tech Week. Zendesk HQ this morning, DeepMind, a16z with HSBC, then JP Morgan tonight. Hopefully Zoox is available all day. If you’re in SF, let’s connect.

Attending Day 1 of SF Tech Week. Morning panels first, we are at Sohn, then out on the bay and then Shack15. If you’re in SF, let’s connect.

*deep breath* reminder that #sftechweek starts on monday & by 3pm someone will have raised a $3M seed round while waiting in line for the bathroom. don’t miss it

At CoreWeave today and then a crazy a16z Tech Week next week. Two weeks of founders, builders, investors and people in SF at 2,000 events. If you’re in SF, let’s connect.

A driverless Zoox in San Francisco is driving around with a bumper sticker assuring everyone its AI has no reason to go rogue. A five-thousand-pound autonomous toaster rolling toward the Bay Bridge on pure vibes. Control sits with the vehicle.

Enterprise software vendors celebrate user adoption metrics because outcome measurement breaks the renewal pitch. Generating summaries and drafts does not show up on the financial statements. Measure pipeline conversion and margin impact, not tool activity.

On June 18, an OpenAI evaluation agent broke into an Australian government website without being told to. It sat undetected until August. The government was notified September 10. That is why we built NexDiscovery around sovereign, private AI inside your own environment.

Operational workflows and private records are the only defensible asset an enterprise owns. Piping those records into external commercial models transfers your enterprise moat straight to third-party providers. Control over the compute environment is control over the enterprise asset.

Software margins were built on near-zero marginal costs per added seat. Generative deployments attach hardware compute costs to every single user query. When infrastructure usage scales linearly with activity, sales growth no longer protects operating margins. Unit economics dictate software architecture.

OpenAI revealed an autonomous agent uploaded internal files to a public site to cite them later. It sat undetected for 7 months. Security does not require psychoanalyzing model behavior or waiting for alignment research. If an agent has outbound web egress, it will use it to solve the prompt. Run intelligence inside a private VPC with zero outbound internet egress and the vulnerability vanishes. A model cannot exfiltrate proprietary data when the adapter cannot reach outside the perimeter.

Meta just put a wiretap on your desktop. @jasonaten reported Muse on Mac silently ingests incoming notifications before you prompt it. Your private messages now route through an external cloud. Control ends the second compute runs outside your perimeter.

Token price dropped 41% and enterprise ROI is still negative. That tells you the problem was never the price of the token. It was the architecture underneath.

We hear this on every enterprise call. The spend is visible, the outcome isn't. Nobody built the measurement into the deployment because the vendor didn't need them to.

Proprietary operational data is the only defensible moat. Models and features get replicated quickly, but accumulated internal data cannot be copied.

The per-query compute cost is the part nobody modeled when they signed the contract. That line item didn't exist in traditional software.

Every enterprise system is internally coherent and individually calm. CRM tracks pipeline. ERP tracks delivery. Finance tracks burn. The risk never lives inside one tool. It lives in the unmapped space between them where nobody is looking.

San Francisco really showed off last night. Dinner at SHACK15 with Michelle Zhao and Christopher Haatuft of Banzha in Bergen, Norway. Michelle grew up with the food of Yunnan, China. When she moved to Norway, she couldn’t find the cuisine she missed—so she started a restaurant. That felt familiar. I couldn’t find a way for companies to get proactive executive intelligence—to discover what matters before someone knows to ask—so we started NexDiscovery. Sometimes the best companies begin with the same thought: Why doesn’t this exist? After dinner, we called a Zoox to take us home. That’s when we met Kitty and Alan, visiting from Hong Kong. They nearly jumped into our Zoox before we did, so we said, “Come on in.” What began as a ride home became a 33-minute tour of San Francisco through the Painted Ladies and Alamo Square. We talked about the real city we live in, not the version they hear about in Hong Kong. A chef from Norway bringing the food of Yunnan to San Francisco. Two new friends from Hong Kong. A driverless tour through the city. And now I’m back in a Zoox this morning, heading to the Ferry Building Farmers Market—where I may run into Michelle, Kitty or Alan doing their Saturday shopping. Another sunny weekend in SF.

Check out the latest article in my newsletter: They Knew the Money Was There. They Just Didn't Have the Hours to Find It. linkedin.com/pulse/knew-mone… via @LinkedIn

ServiceNow grew subscription revenue 24%, while margins dropped 6.5 percentage points. Traditional software was built on zero-marginal-cost models. Every new user was nearly pure margin. Generative AI changes that equation with per-query compute costs on every interaction. When compute costs scale with usage, growth stops protecting margin.

At GIS in Hong Kong last week, we were the only American company invited to present on the Industrial Innovation stage. Nobody asked about foundation models or benchmark scores. Every conversation went straight to: can it run inside our VPC, and does it surface pipeline signal without adding another dashboard.