today i'm joining @lightspeedvp as Chief Marketing Officer.
after doing thousands of IPOs at Nasdaq and building Redpoint's exceptional platform team, i've learned:
it's not just one launch video. it's the relentless drumbeat that get you heard.
and that's what i'm here to build for Lightspeed and our founders.
let's make some noise!
ai video was the warmup. the real game is worlds you can step into and train in.
@drivascos from @GoogleDeepMind walked @clairejyz and me through genie 4 and what real-time world models mean for robotics and simulation. full ep 👇
AI is moving beyond generating content and into simulating entire worlds.
@drivascos, Group Product Manager at @GoogleDeepMind, joins @clairejyz and @Machiz on Lightwork to discuss Genie 4 and what real-time world models could mean for the future of robotics, simulation, and even entertainment.
Diego explains how Genie 4 combines video generation, contextual memory, and real-time audio to create simulated 3D environments that users can navigate and modify. It works alongside Gemini, which serves in a “director”-like role.
“Anyone that has an idea effectively has the medium to turn it into this interactive concept, a more immersive concept that you can showcase and maybe debate or iterate on,” says Diego.
These worlds could let filmmakers explore new storylines and robotics researchers test scenarios that are hard to recreate in real life.
Full episode below.
0:00 Genie 4 and World Models
0:58 World Models vs. Video Models
1:52 Genie’s Biggest Research Milestones
3:09 What Enabled Genie 4’s Biggest Leap?
4:33 Gemini’s Role in Directing World Models
5:29 How Do Gemini and Genie Work Together?
7:26 How Does Real-Time Audio Generation Work?
9:10 What Makes Real-Time Audio So Difficult?
10:08 How Are Creators Using Genie 4?
12:57 Genie’s Real-World Applications
15:12 Personalized World Experiences for Consumers
16:34 Genie 4 Being Used in Robotics
18:27 Hybrid World Models
19:32 Where Do World Models Still Fall Short?
20:26 Where World Model Simulations Fall Apart
22:17 Genie 4’s Path to Wider Availability
23:13 Most Unique Genie 4 Use Cases
23:52 Scaling World Models Amid Compute Costs
24:48 What's Next for Genie and Project Genie?
compute deals happen on WhatsApp and Slack. the team behind Compute Desk says that won’t scale.
@andrawesbahou sits down with @jmover and @katkzhang on The Investment Memo.
thanks to andy for sharing what the team is building, its a must watch episode.
Compute is becoming a financial asset.
@ComputeDesk co-founder and CEO @andrawesbahou joins Lightspeed partners @jmover and @katkzhang on The Investment Memo to discuss how that market is taking shape.
Andy argues that reliable delivery of compute brings buyers and sellers into the market. Their transactions help establish the price benchmark others can use to trade and manage risk.
Tune in to the full conversation below.
0:00 How Compute Became A Financial Asset
3:08 Historical Parallels to Oil
5:57 What Compute Desk Builds
7:37 Why Capital Markets Have Not Coalesced Around Compute
9:07 Who Could Benefit From The Financialization of Compute?
11:44 How Compute Volume Can Grow
14:46 The Three Products: Compute Signal, Trader And Clear
19:25 Why The Reference Rate Is "The Ultimate Prize"
20:49 Building A Team With Dual DNA
27:53 Why Physical Settlement Matters
34:18 The Paper Versus Physical Market
37:54 The Paradox Of Market Efficiency
39:29 The Fundraising Process And The Ten Year Vision
Compute is moving beyond IT budgets and into capital markets.
We’re leading @ComputeDesk’s $45M Seed round as the team builds infrastructure to support that shift.
Co-founders Andrawes (Andy) Bahou (@andrawesbahou) and David López Mateos (@SenorScience) bring experience in the infrastructure and financial markets behind this opportunity. Andy led GPU build-outs at Meta and worked in derivatives pricing; David’s background spans commodities and pricing algorithms.
That experience matters when compute commitments can stretch years into the future. Buyers and sellers need ways to price those commitments, manage risk and verify that contracted capacity is delivered.
Compute Desk is building across the whole stack, from pricing benchmarks and trade execution to physical settlement.
Congratulations to Andy, David and the entire team!
CC: @jmover, @katkzhang
Nearly every chip on the market forces a tradeoff between capacity and speed.
If you want a chip that can run tasks fast, handle long context, and serve multiple users concurrently at hyperscale, we're going to need a revolutionary new architecture.
@goodwin_ml and the @fractile_ai team are taking a full-stack approach to tackle this wildly difficult problem. Had a great time catching up with the team to talk through our investment in Fractile.
BREAKING: FRACTILE RAISES A $720M SERIES C CO-LED BY LIGHTSPEED
I got an inside look at the technical breakthroughs @fractile_ai has been quietly building for years, with Founder and CEO @goodwin_ml, alongside @guruchahal and @jonahcader
🧠 Why memory, not compute, is AI’s real bottleneck - and why today’s chips force a choice between big but slow (HBM) or fast but small (SRAM)
🧱 Stacking DRAM directly on logic gives ~20x the bandwidth per chip of HBM and 10x+ the capacity of SRAM
🔁 Why they chose not to bake the transformer into silicon when models change faster than a chip’s 3-5 year life in the data center
🌍 How a team spanning the UK, US and Taiwan pulled off a “secret skunkworks” in advanced packaging
🧑💻 Hiring across silicon, systems and software in London, Bristol and Palo Alto
(My favorite part was hearing how the team flew to London, drove three hours to Bristol for one day with the Fractile team, and even got to catch an eclipse! 🌘)
FULL SEGMENT BELOW
@lightspeedvp
The next leap in intelligence could depend on advances in compute.
We’re excited to co-lead @fractile_ai’s $720M Series C as they build an inference accelerator to support the next generation of models.
Reasoning and agentic workloads are expanding demand for inference, making the speed and cost of running frontier models increasingly important.
The team is building a chip-first frontier AI company, rethinking accelerator architecture from the ground up. Its systems are designed to run frontier models significantly faster and at a fraction of the cost.
Congratulations to the entire Fractile team!
CC: @guruchahal, @jonahcader, @jshen99
If we want to meet the demand for frontier tokens, we’re going to need a massive buildout of the underlying compute infrastructure.
Long-horizon tasks, heavy reasoning workloads, and subagents are all accelerating the need for new processors.
@goodwin_ml has brought together an incredible team, creating hardware to push AI’s frontier forward. Today, he and the team are building chips and systems that can deliver both speed and scale. Proud that we @Lightspeedvp are co-leading @fractile_ai’s Series C!
CC: @guruchahal, @jshen99
We’re delighted to announce that Fractile has raised $720M in a Series C led by @thrivecapital@foundersfund@lightspeedvp and @redpoint, with further participation by brilliant new and existing investors.
pov: you’re testing a boat when the motor cuts out in the middle of a lake.
you need to figure out what happened, whether you can get it running again, and if you can keep testing.
@brycestrauss uses this example to explain how engineers investigate a failure.
with their test data in @Nominal_io, they can compare current readings with earlier tests and notes from the team.
Bryce sat down with @clairejyz and me on Lightwork to discuss how Nominal has evolved to help hardware engineers and what the future of physical AI looks like.
How can AI help engineers build hardware we can rely on?
@brycestrauss, co-founder of @Nominal_io, joined @clairejyz on Lightwork to discuss how the team is approaching that question.
Nominal is an intelligence platform for hardware engineering, connecting hardware data with engineering context to help teams build, test, and operate complex machines.
For Bryce, the opportunity in physical AI is helping engineers build more ambitious systems, from faster aircraft to new energy technology.
Bringing those systems into everyday use means scaling production with confidence that the hardware will perform as intended.
Full episode below.
0:00 Welcome To Light Work
1:29 What Nominal Builds For Hardware Teams And Origin Story
5:38 Series C And The Analyst Agent
6:40 The Boat Example And Why Context Matters
10:58 Data Infrastructure And The Three Step Process
12:59 The Pitch To Aerospace And Defense Executives
15:58 Team And Hiring Philosophy
17:59 The Future Of Physical AI And Reindustrialization
23:54 The Next Decade And Thank You To Physical Engineers
AI in Space
Google says it received the first-ever AI response from a data center-class TPU in orbit. It asked Gemma 4 to write a haiku about space. The satellite sent back this response over Norway:
“Blue marble below / Data streams in silent grace / Starlight feeds my core.”
More on Lightwork this week:
→ The $6.3B Binance Hack
The FBI attributed the roughly $6.3B Binance theft to North Korea’s state-run hacker unit, TraderTraitor. Binance says attackers used a fake job interview to get malware onto an engineer’s work computer, then used an internal AI tool to reach other engineers’ systems.
→ Cybersecurity Gets Physical
Cyberattacks targeted supermarket refrigeration systems across Europe and plasma storage sites in France, showing how digital vulnerabilities can put food and medical supplies at risk. No blood products were lost in the plasma-storage incident.
→ Is the Internet Dying?
Hosts @clairejyz and @Machiz discuss AI slop and how AI overviews are affecting Reddit’s traffic. Andrej Karpathy’s HackerTube takes a different approach to discovery, surfacing YouTube videos recommended on Hacker News. As content gets easier to produce, does curation become more valuable?
0:00 Welcome to Lightwork
1:09 Google's AI Breakthrough in Space
1:45 Project Star Catcher and Space Data Centers
6:01 Google's Next Steps for Space Data Centers
7:27 Binance's $6.3 Billion Crypto Theft
10:56 Cyberattacks Target Supermarket Freezers
14:27 Is the Internet Dying?
15:57 Andrej Karpathy's Answer to AI Slop
16:58 Reddit Loses Traffic to Google's AI Search
20:02 AI Reviews vs. Human Commentary
21:40 How Marketers Can Leverage Reddit in the AI Era
23:31 Why Content Curation Matters in the AI Era
The speed at which AI agents can infiltrate an environment demands a fundamental change in how security teams detect and respond to attacks. CISOs need to rethink how quickly their teams can identify a threat, investigate it and take action. There's a significant opportunity for tools that shorten the time from detection to containment
More in our CISO report for this year’s @FortuneMagazine Cyber 60: lsvp.com/cyber60-2026-2027/
How does a corporate card startup become a $60B AI product factory?
In our first on-camera interview a decade in the making, Ramp co-CEO Eric Glyman (@eglyman) answers just that question.
Today, @tryramp is best known for its corporate cards, or maybe its marketing stunts with the likes of Kevin from ‘The Office,’ Tony Hawk and a 70-minute Broadway musical about paying bills.
But Eric says his $60B startup, NYC's most valuable, feels less like a fintech and more like a productivity company driven by one simple question: If a business spends $100 to make $200, how can Ramp improve that ratio?
Nearly 100,000 companies use Ramp today for everything from expenses to managing AI spend. But Eric says that when he zooms out to the millions who still aren't customers, Ramp feels painfully early still.
On The Upstarts Podcast, Eric shares why he’d rather run Ramp like Amazon than Google; why Ramp hires for proof of work, not credentials, from a 14-year-old intern to a 28-year-old CTO; and how being the first customer for Anthropic’s Claude Code and Cognition’s Devin led Ramp to build its own AI tools.
Plus: The real story of Paribus, his first startup with co-founder Karim Atiyeh (@karimatiyeh); what they learned at Capital One; and a warning for Instinct and Muse.
This season is presented by @Atlassian and @chargebee 🫡
CHAPTERS
00:00 Introduction
01:51 What Ramp Actually Does
03:27 From Gym Charges To 1% Of U.S. Corporate Payments
05:13 Glass, Inspect And Building AI Tools For Itself
07:45 An AI Sales Assistant Before ChatGPT
09:45 How Ramp Decides What Products To Ship
11:14 First Customer For Claude Code And Cognition’s Devin
13:09 Why Ramp Wanted To Build Like Amazon, Not Google
15:42 Hiring For Proof Of Work: The 14-Year-Old Intern
18:43 A .300 Batting Average And ‘Growth Without Fear’
21:17 What Ramp’s Founders Learned At Capital One
24:00 The $1,000 Flight Refund Behind Paribus
26:34 Bots, A Cease-And-Desist, And A Lesson For Instinct And Muse
32:00 Marketing 101: Bill Pay The Musical And Kevin In A Glass Cube
36:01 Business Books And Helping People Do Their Life’s Work
39:03 How A 28-Year-Old CTO Catalyzed A Co-CEO Change
43:55 ‘Still So Early’: How Ramp’s Founders Stay Hungry
55:26 John Wooden And Why Ramp Competes Against Itself
Thanks for watching!
What I'm seeing across the industry right now is that the biggest bottleneck to AI adoption isn't model capabilities. It's data infrastructure.
Today, we're expanding the Eon Data Foundation with the launch of Eon Yukon, a new serverless data engine purposely built to change that.
AI agents need fresh, contextualized data to act.
Specialized models need a continuous feed of data and feedback to improve. Both are still served by a stack, an architecture, and methods built for dashboards:
-Spark and Flink clusters sized for peak, plus compaction jobs to keep tables usable.
-A pipeline and a copy for every consumer, staged bronze to silver to gold.
-Hand-written jobs, wired together by an orchestrator and kept alive by specialists.
The result: shipping data is complicated, risky, and expensive. While coding agents ship apps in hours, they rarely build or run data pipelines. The data is too volatile, and a mistake costs too much.
With Yukon, data moves from ingestion to processing to serving in one continuous flow, without the infrastructure and pipelines traditionally required to connect them.
Events become compacted, sorted, deduplicated Iceberg tables within seconds.
Real-time aggregations run inline, with fully managed state. Agents and applications can access fresh data with sub-millisecond latency through a Redis-compatible API or through our SQL and spark compatible query engine.
Beyond moving and serving data, Yukon makes it continuously understandable - Live column-level statistics, accessible through MCP, give agents visibility into the data. Built-in monitoring, sampling, and data minimization help teams maintain control.
At 1M events/sec and equal freshness, Yukon used 4x less compute than Spark and 7x less than Flink. The benchmark is public: run it yourself with Eon’s ingest-bench.
This is another major step toward our vision of an Autonomous Data Foundation for the AI era, built to both protect enterprise data and unlock its full potential.
Coming tmrw, an interview a decade in the making 👀
I've known Ramp co-CEOs Eric Glyman (@eglyman) and Karim Atiyeh (@karimatiyeh) since college. I was an early user of their first startup, Paribus.
Now, all these years later, Eric joins The Upstarts Podcast. Excited to share!