Forget Nvidia. Farzad Says SpaceX Is the New King of AI

Everyone still points at Nvidia as the king of AI. Farzad’s Sept 2, 2026 video says that framing is already wrong. The company that will matter more, and that Nvidia’s long-term demand depends on, is SpaceX. The opening proof is a model. SpaceXAI’s Grok 4.6 scored about 61 on the Artificial Analysis Intelligence Index, level with OpenAI’s GPT-5.6 Sol and one to two points behind Anthropic’s Claude Fable 5 and Claude Opus 5. Price is the buyer’s tell. Grok lists at $2 per million input tokens and $6 per million output. Sol, the cheaper near peer Farzad names, lists at $5 and $30. On long multi-step knowledge work, he says Grok reached Fable-level results while using about half as many turns and roughly one quarter of the input tokens Opus 5 Max burned. Frontier intelligence at one of the best cost profiles in the market is half the thesis. The other half is why a rocket company closed that gap this fast, and why the model may be the least important piece.

Grok 4.6 and the Grok Bot harness

A chat box that answers a question is useful. An agent that opens tools, finds files, finishes the work, checks the result, and returns with something done is the product Farzad cares about. Grok Bot is SpaceX’s agentic harness competing with OpenClaw, Hermes, Claude Code, and Codex. It gives the agent a persistent cloud computer, browser, terminal, filesystem, memory, and the user’s tools. Assign a job, close the laptop, let it keep working. Multiple bots can pass work and pull a human in when judgment is required.

Farzad’s uses (email triage, farzad.fm, a farzad.money book he says was up versus the S&P 500 at recording time, YouTube research, Discord, SEO, and Shorts) are his anecdotes, not SpaceXAI metrics. From Grok 4.5 to Grok 4.6 he cites a five-point Intelligence Index gain in about a month at the same price, with stronger agentic work. SpaceXAI still knows how to train a frontier model. The next advantage is joining the model, the harness, the compute, and the product team in one feedback loop.

SpaceX closed its acquisition of Cursor’s parent, Anysphere, on Aug 14, 2026. Cursor joins the SpaceXAI team with experience shipping agent products developers already use, plus data from real agent workflows. Improve the model for the harness. Improve the harness for the model. That loop is hard to rent from the outside.

Colossus speed and the physical loop

Training still means a giant room of computers doing absurd amounts of math. A large AI data center is power, cooling, heat, networking, and schedule, not just racks of GPUs. Farzad cites SpaceXAI’s claim that the first Colossus cluster came online in 122 days inside an old factory shell. Musk posted that 122-day figure on Sep 2, 2024. SpaceXAI still says it then doubled the original cluster in 92 days. Reporting on Colossus 2 / Macrohard puts the other side’s first cluster at about 91 days. SpaceX’s IPO language said Colossus and Colossus II together were about 1.0 GW of compute. That is the S-1 print, not a Sept 2, 2026 operating meter. On the Aug 4, 2026 Q2 call the company printed 1.4 GW of nameplate compute company-wide and said it expects more than 2 GW by year-end 2026. The average data center take he uses for comparison is around two years. That is roughly five times faster on his framing.

Tesla shows a related playbook in Austin. Cortex 1 has more than 100,000 H100-equivalent GPUs in production. Cortex 2 has started training workloads and is ramping toward more than 130,000 H100 equivalents. Tesla and SpaceX are separate companies. The shared lesson is culture: pull power, cooling, networking, software, and construction into one room and attack the schedule as one system.

The compounding loop is simple. More compute means more experiments. More experiments improve the model. A better model attracts users and revenue. That demand pays for more compute. The next model arrives faster and cheaper. On SpaceX’s Aug 4, 2026 Q2 call, Musk said the company ended June with 1.4 GW of nameplate compute, expects more than 2 GW by year-end 2026, and that cumulative compute by the end of 2027 may be “closer to 10 gigawatts of compute than five.” Farzad compresses that to a 5x. That 5x is his read, not a word SpaceX used. Treat 10 GW-class compute by end of 2027 as the company target, still a plan, not a finished print.

Orbital compute and Starship mass

Earth AI still hits electricity, land, cooling, permits, and substations. Political backlash can slow the rollout further. SpaceX’s bet is sun-synchronous solar, Starlink laser links, and Starship launch mass.

A satellite in the right orbit can see near-constant sunlight. Solar turns that into electricity for the computers onboard. Starlink lasers move data between satellites and can send results back to Earth. Sunlight alone does not make orbital compute cheap. The limit is how much solar area, compute, cooling, and satellite mass you can manufacture and launch. Starship changes that mass equation. Farzad frames full reuse as putting well over a thousand times more mass per year into orbit than current rockets.

SpaceX’s S-1 says orbital AI sats as early as 2028, early sats around 100 kW of compute, and a long-term goal of about 100 GW of new orbital compute per year. Those are company targets, not a fleet on orbit. Farzad’s “about 2.5 times US data center capacity being built in 2026” is his comparison. The S-1’s own comparator is about one-fifth of 2025 US electricity production. That would take thousands of Starship launches a year, larger solar arrays, radiation-tolerant chips, huge radiators, and a manufacturing system that replaces older sats as new chips arrive. If the economics work, SpaceX is the AI customer, data center operator, satellite maker, network, and launch provider in one package. No other AI company can buy that stack off the shelf today.

Custom chips and Terafab

Frontier labs still lean on Nvidia, TSMC, memory suppliers, packaging houses, and a supply chain that sells to everyone. When chips are scarce, access, price, and delivery time become the limit.

Tesla has spent years on custom inference chips for cars that must read camera video, understand the road, and act inside a tiny power budget. The company is working through AI5 and AI6, with materials targeting production in 2027 and 2028. Elon has described AI7, AI8, and AI9 on a faster design cycle. Terafab pushes further. SpaceX, Tesla, and Intel have started work on a chip facility meant to put design, logic and memory fabrication, and advanced packaging inside one closed-loop plant. Farzad compares the ambition to Gigafactory Nevada matching world battery capacity from one site. The point of the closed loop is cycle time: design, test, find the bug, change it, run again without shipping the problem across several companies and countries.

Terafab plans two broad chip families in his telling: one tuned for vehicles and Optimus robots on Earth, another built for radiation, power, and heat limits in orbit. If it works, the model team that ships Grok 4.6 or Tesla’s FSD stack can tell the chip team which operations burn time and energy. The chip team changes the silicon. The data center team changes power and cooling. The satellite team changes the solar array and radiator. The next training run sits on hardware designed for the actual workload. That is how Tesla approaches autonomy across model, software, onboard computer, cameras, and car. Bring that loop into Grok, Optimus, orbital compute, and FSD, and the cost of useful intelligence can fall again.

Musk culture and the flywheel

Elon’s pattern, as Farzad retells the five-step process, is question the requirements, delete unnecessary steps, simplify what remains, speed up the cycle, then automate it. That process creates mistakes, impossible schedules, chaos, and missed dates. When it works, it changes an industry’s cost curve.

The AI version is visible in this video’s stack. Build the data center faster. Train models faster and cheaper. Put the model inside an agent that does real work. Design the chip around that workload. Manufacture the chip closer to the team. Put compute next to abundant solar in space using the rocket and satellite network you already own. Put the resulting intelligence into products that already have users. Then use AI and robots to improve the whole system that produces more AI and more robots.

Better models help design chips, software, factories, and satellites. Better chips cut the cost of training and running models. Cheaper intelligence makes robots more useful. Useful robots build more cars, data centers, chips, and satellites. More infrastructure creates more intelligence. The cycle starts again higher up. Farzad’s book Master Plan is the longer cut of how Tesla, SpaceX, SpaceXAI, X, Optimus, energy, and chips fit together. This Exclusive is the short cut: Nvidia sells the picks and shovels. SpaceX is trying to own the mine, the factory, the power plant, the launch pad, and the agents that spend the tokens.

Keep the hedges. Artificial Analysis scores and list prices are measured. Farzad’s peer set is the Aug 12, 2026 snapshot. On Sep 1, 2026, Claude Fable 5.1 (max) scored 66 and took the lead. He does not cite 5.1. Colossus 122-day and 91-day builds are company-side speed claims. The ~1 GW Colossus-plus-Colossus-II figure is the S-1 print, not a Sept 2, 2026 meter. Farzad’s 5x by end of 2027 is his read of Musk saying closer to 10 GW than 5. Early 100 kW orbital sats in 2028, 100 GW per year of new orbital compute, and Terafab’s multiple-of-global-capacity ambition are targets and pitches, not finished capacity. Cursor closed Aug 14, 2026 (SpaceX 8-K). The thesis is not that Grok is forever the smartest model. The thesis is that the company that owns model, agent, compute, chips, launch, and constellation can keep winning even when the scoreboard rotates.

This Exclusive is from the long-form at https://www.youtube.com/watch?v=d6_JGxriQ5w.