The Two Harvard Dropouts Who raised $800M to take on NVIDIA

21 insights saved from this video by @technology

@technology
Advanced AI will look less like a machine and more like a giant distributed brain spanning countless chips in one data center.
@technology
A hidden analog clock-domain failure produced incorrect attention results and escaped a massive digital FPGA verification cluster.
@technology
Fixing the chip required aligning two clock signals within 50 picoseconds on every chip, two billion times per second.
@technology
Pre-fetching cut the time from receiving silicon to running inference from a competitor’s 10 months to 40 days.
@technology
Inter-chip communication on current NVIDIA products takes roughly 4,000 nanoseconds, against a physical lower bound near two or three.
@technology
The first generation of this low-voltage inference technology runs at less than half the voltage of any other AI chip.
@technology
A custom interconnect stack above Ethernet’s second layer cuts chip-to-chip latency more than fivefold.
@technology
TSMC funded a manufacturing experiment, confirmed its yield improvement, then applied the change across the production line.
@technology
High-frequency-trading engineers were the first believers in kernels-first programming, because they already bypassed compilers.
@technology
Project-based recruiting maps the hardest technical problems, finds who solved them, and tracks those people through repeated conversations.
@technology
The strongest hardware teams combine seasoned scale experts with young builders whose chips-on-their-shoulder drive creates extraordinary execution energy.
@technology
High-speed AI decoding depends primarily on moving data, not solving harder mathematics, across and within chips.
@technology
Aggressive spending makes sense when every day without shipping leaves substantial opportunity on the table.
@technology
Aggressive fundraising got the rack to market quickly. Existing investors later doubled or tripled down.
@technology
Repeated contrarian bets get harder to defend emotionally as the stakes grow larger.
@technology
AI hardware requires funding the chip, boards, interconnects, cooling, networking, and the entire cluster.
@technology
Pre-fetching turns waiting time into execution by finishing every independent task before the missing component arrives.
@technology
The 40-day result came from making everything possible routine before the chip arrived, leaving only final assembly and minor adjustments.
@technology
Long-tenured engineering teams draw confidence from hard-won experience that can overwhelm newcomers facing the same problems.
@technology
Fast local communication cannot rescue an AI system when communication between chips remains slow.
@technology
Building a chip better than every previous AI chip requires enough naivety to challenge what experts accept.

More from Invest Like The Best

See everything from Invest Like The Best

Post on X

Keep what you found in The Two Harvard Dropouts Who raised $800M to take on NVIDIA

A free account saves any of these cards to your Collection, and Korva resurfaces them so you actually remember.