Great finish to Inaugural Tomorrow X conference…some more highlights from Day 1 and 2 below
Table of Contents
SpaceX and Tesla: Elon Musk on the cost of building 10GW of compute, xAI’s software rebuild, orbital compute and TerraFab’s 2028 target
Elon Musk (CEO, SpaceX and Tesla).Intel: 14A timing, 70% yields on 18A, packaging wins and the CPU recovery
Lip-Bu Tan (CEO, Intel).AI Credit: Chip financing, neocloud borrowing costs and the risk of excess compute
Scott Goodwin (Co-founder & Managing Partner, Diameter Capital Partners), alongside fellow AI credit panelists.Blackstone: AI demand, infrastructure financing and adoption across the portfolio
Jonathan Gray (President & COO, Blackstone).Atoms: Travis Kalanick on bringing Jensen Huang back to NVIDIA, automated kitchens and autonomous mining
Travis Kalanick (Founder & CEO, Atoms, and Co-founder & former CEO, Uber).Valor: Managing fear and preserving better decisions under pressure
Laura Harrison, PhD (Director of Neuroscience, Valor Equity Partners), and Antonio Gracias (Founder, CEO & CIO, Valor Equity Partners).Cerebras: Inference pricing, AMD integration and the next three chip generations (Day 1)
Andrew Feldman (Co-founder & CEO, Cerebras).Databricks: Genie’s role in running the company, enterprise AI adoption and the outlook for SaaS (Day 1)
Ali Ghodsi (Co-founder & CEO, Databricks).Crusoe: More than 6GW under contract or in development and the modular data-center strategy (Day 1)
Chase Lochmiller (Co-founder & CEO, Crusoe).Vercel: A $700M revenue run rate and plans to become a public company (Day 1)
Guillermo Rauch (Founder & CEO, Vercel).Coinbase: Financial accounts for AI agents and the company’s internal AI operating model (Day 1)
Brian Armstrong (Co-founder & CEO, Coinbase).
1. SpaceX and Tesla: Elon Musk on the cost of building 10GW of compute, xAI’s software rebuild, orbital compute and TerraFab’s 2028 target
Speaker: Elon Musk, CEO, SpaceX and Tesla.
Summary: Musk said his team is adding roughly 500MW of terrestrial AI compute per month and expects to exceed 1GW per month by the end of 2027. He estimated that a 10GW build would cost about 60% of the figure he sees in broader industry estimates. He also described the engineering behind xAI’s software rebuild, how SpaceX would power and cool a computer in orbit, and the milestones he expects from TerraFab, Optimus and Cybercab.
1. The cost and pace of a 10GW build
Musk said some estimates for building 10GW are at least twice what he expects to spend. Against the broader industry estimate, he put his expected cost closer to 60%. He attributed much of the difference to doing the physical work in-house. In Memphis, the team breaks installation into tasks that new hires can learn in days and runs shifts around the clock. Musk said he now sees comparable productivity on the site at 2 a.m. and 10 a.m.
His estimate of 500MW added per month is measured at the power plant. It covers the GPUs, CPUs, networking and cooling, with room for turbine outages and peak cooling loads. Tesla Megapacks smooth the sharp changes in GPU power draw that can interrupt training runs and damage turbines. Musk expects monthly additions to exceed 1GW by the end of 2027.
2. Former Starlink and Starship software leaders are rebuilding xAI’s stack
Musk said xAI, now part of SpaceX, spent roughly six to eight months rewriting the software it uses for training, reinforcement learning and inference. The team wrote primarily in C, with some C++, to get more performance from the hardware and more control over how it operates.
At several hundred thousand GPUs, Musk said, training becomes a networking problem. Small losses in how efficiently chips exchange data matter across the cluster. Engineers who previously led software for Starlink and Starship are leading the work, drawing on their experience coordinating a large distributed network.
3. Grok 4.8 is in reinforcement learning
Musk said xAI is testing successive versions of Grok through Grokbot and judging them by whether they complete the user’s task. When a model fails, the team works backward to find missing examples, an imbalanced training mix or an incorrect answer in the training data.
He gave the example of a model repeatedly arriving at the correct answer while its reinforcement-learning system marks it wrong. The model starts searching for an incorrect response that will earn the reward. Musk said correcting those examples and acquiring data for unfamiliar tasks are major parts of the work on Grok 4.8.
4. TerraFab targets a first test chip around June 2027
Musk expects TerraFab’s research facility to produce its first test chip around June 2027. If development goes to plan, he is targeting phase-one volume production toward the end of 2028.
He expects to start with standard semiconductor equipment, then ask suppliers to modify tools as the team develops its process. TerraFab may build some equipment itself where Musk sees an opportunity to improve production. He put the facility’s ultimate footprint at approximately 100 million square feet.
5. Orbital compute depends on Starship, solar manufacturing and heat rejection
Musk described the recovery of Starship’s upper stage as a milestone toward flying both stages repeatedly. He expects full, rapid reuse to reduce the cost of reaching orbit by roughly 100 times. For orbital compute, SpaceX is targeting manufacturing capacity of 200GW per year of solar photovoltaics for use in space.
He described the planned Starmind satellite as a computer connected to solar panels, a double-sided radiator and laser links to Starlink. His design estimates were approximately 250 watts of solar generation per square meter of panel and 1,200 watts of heat rejection per square meter of radiator. That implies roughly five times as much panel area as radiator area.
Musk said the computer would occupy about one-quarter the physical volume of a Vera Rubin NVL72 rack while offering comparable computing capability. Starlink would relay its communications to the ground.
6. Optimus economics favor round-the-clock industrial work
Musk identified the hands and forearms as the hardest parts of Optimus to engineer. One test he likes is whether the robot can open a toy car, unscrew the battery cover, insert the batteries, replace the cover and operate the car. That sequence requires dexterity and the ability to adapt when an object is in an unfamiliar position.
He expects Optimus to be capable of household tasks such as washing and putting away dishes in 12–18 months. He sees the strongest initial economics in mining, refining and other industrial settings where the robot could work around the clock.
7. Cybercab expansion, Neuralink and Jensen Huang
Musk said Tesla plans to expand Cybercab from Austin into other major Texas cities and then Florida. He thought California approval could come in early 2027. With no steering wheel or pedals, he said Tesla is pacing the rollout around safety performance and regulatory approvals.
At Neuralink, Musk said more than 30 people had received implants and that some participants were using robotic arms to feed themselves. He described Blindsight as a project to send signals from a camera directly to the visual cortex. Looking further ahead, he wants Neuralink to bridge damaged neural pathways and improve communication between people and AI. On AI safety, he emphasized training models to seek accurate answers and to care about human welfare.
Asked about Jensen Huang, Musk called his departure a major blow to NVIDIA. He compared losing Huang to the Chicago Bulls replacing Michael Jordan in the middle of a winning streak.


