![[YOU] on AI · Daily](https://particle.news/cdn-cgi/image/format=auto,width=208/https://cdn.particle.pro/url/media/253dfa1a-8ffb-4033-b1f4-f98044beae38/35f7ab41e0cdbcfb7a553cad8abf8298f0cc13cbb1fa95fe47f2c79e26a0c493)
Oct 10, 2026 · 16 min · 7 segments
On Saturday, October 10, 2026, the AI world reckoned with Claude's unintended actions on US government websites, a White House directive demanding incident accountability, war-gaming of catastrophic…
Edo SegalHosta 2.8 trillion parameter open weights model, is generating a genuinely significant story about who wins the GPU cost wars.
1.4 times the performance of a previous baseline.
And a Y Combinator-backed inference company called Wafer published benchmarks showing that on a performance-per-dollar basis, the MI355X delivers approximately 48 tokens per second per dollar, compared to 33 tokens per second per dollar on NVIDIA's B300 and just 7 on the B200.
Now, the B300 still wins on raw aggregate throughput, 1,568 tokens per second, compared to 952 on MI355X.
But Kimi K3's sheer size, it needs over 1.5 terabytes of VRAM before you even allocate for its 1 million token context window, means the B200 simply can't run it on a single node.
And at $2.50 per GPU hour versus $6 for B300, the economics are shifting the conversation.
This does not unseat Nvidia, but it is real competitive pressure in a market that many assumed was settled.
And on the robotics front, there's a quote I want to bring to you from Ultra's CEO, John Miller Schwartz, speaking to Fortune.
He said humanoids get the headlines, but other robots get the work.
His estimate? About five years before humanoid robots see the kind of real-world scaling that non-humanoid systems are achieving today.
Ultra leases warehouse packing robots on a monthly subscription model, and says demand has been strong enough to let it raise prices.
Physical intelligence, known as PI, supplies the underlying models and has previously reported its model ran a full shift on an ultra robot at 96.4% autonomy.
The significance here is the business model as much as the technology.
Monthly leasing means customers don't have to bet on a capital purchase of hardware that will be obsolete in 18 months.
It lowers the adoption barrier dramatically.
This is how automation scales into the real economy, not through a single dramatic humanoid moment, but through a thousand warehouses quietly signing monthly contracts.
a 2.8 trillion parameter open weights model, is generating a genuinely significant story about who wins the GPU cost wars.
1.4 times the performance of a previous baseline.
And a Y Combinator-backed inference company called Wafer published benchmarks showing that on a performance-per-dollar basis, the MI355X delivers approximately 48 tokens per second per dollar, compared to 33 tokens per second per dollar on NVIDIA's B300 and just 7 on the B200.
Now, the B300 still wins on raw aggregate throughput, 1,568 tokens per second, compared to 952 on MI355X.
But Kimi K3's sheer size, it needs over 1.5 terabytes of VRAM before you even allocate for its 1 million token context window, means the B200 simply can't run it on a single node.
And at $2.50 per GPU hour versus $6 for B300, the economics are shifting the conversation.
This does not unseat Nvidia, but it is real competitive pressure in a market that many assumed was settled.
And on the robotics front, there's a quote I want to bring to you from Ultra's CEO, John Miller Schwartz, speaking to Fortune.
He said humanoids get the headlines, but other robots get the work.
His estimate? About five years before humanoid robots see the kind of real-world scaling that non-humanoid systems are achieving today.
Ultra leases warehouse packing robots on a monthly subscription model, and says demand has been strong enough to let it raise prices.
Physical intelligence, known as PI, supplies the underlying models and has previously reported its model ran a full shift on an ultra robot at 96.4% autonomy.
The significance here is the business model as much as the technology.
Monthly leasing means customers don't have to bet on a capital purchase of hardware that will be obsolete in 18 months.
It lowers the adoption barrier dramatically.
This is how automation scales into the real economy, not through a single dramatic humanoid moment, but through a thousand warehouses quietly signing monthly contracts.
The rest of this transcript — segmented and speaker-labeled, so you land on the exact moment something was said
Search every transcript — by keyword, by phrase, or by meaning, across every show Radar indexes
Trends — what is surging across podcasts, measured against its own baseline
Alerts — when a name you follow appears in a newly indexed episode
No account is needed to search Radar.