Permanent copy of issue 47. Today's issue is always the front page.
Google released Gemini 3.8 Live and Gemini 3.8 Live Extended Thinking on Tuesday, putting real-time voice agents on its API at $0.005 a minute of audio input. Salesforce and Nvidia said Koa, a CRM reasoning model built on open Nemotron weights, will run inside Salesforce's own systems. Nvidia said a Santa Clara factory cut power from four megawatts to three on a utility signal. Periodic Labs said a trillion-parameter model trained on 1,300 of its own chips beat larger lab models at reading X-ray experiments.
Date: 2026-09-16
Updated: 2026-09-16 06:18
Status: LIVE
Broadcast
Google priced a live voice at half a cent a minute of audio. That is a cluster product sold by the minute, not a model a household can keep. Salesforce said Koa will run on its own machines and keep customer files inside its walls; the training started from Nvidia's open Nemotron weights, but the meter is Salesforce's. Nvidia said Silicon Valley Power has already sent more than 200 demand signals to one factory, cutting load from four megawatts to three without stopping the high-priority jobs. Periodic Labs said 1,300 of its H200 chips were enough to beat GPT-6 Astra at reading X-ray patterns from its own benches. A specialist lab with its own data is not the same as a rented hall, and it is not a laptop.
Watch
Stories
Google released Gemini 3.8 Live and Gemini 3.8 Live Extended Thinking on Tuesday for the Gemini API and Google AI Studio, and said the extended model took first place on Artificial Analysis's Speech to Speech Quality Index at 82.6. Google priced the models at $0.005 a minute of audio input and $0.018 a minute of audio output, and said they switch among 97 languages and keep talking while they call tools in the background.
-- A household that talks to a phone is still paying Google by the minute; the model does not sit on the device.
2026-09-15 / Google / chip / central / #google-ships-gemini-3-8-live
Salesforce and Nvidia said Tuesday at Dreamforce that Koa, Salesforce's first CRM reasoning model, was built by post-training Nvidia's open Nemotron 3 Super on a synthetic set drawn from 27 years of CRM work. Salesforce said it controls the weights, runs the model on its own systems, used no customer data, and will put Koa into customer pilots in October, with general availability expected in U.S. regions in winter 2026.
-- An open-weight starting point that never leaves Salesforce's hall is a company boundary, not a desk the household owns.
2026-09-15 / Nvidia / chip / mixed / #salesforce-and-nvidia-launch-koa
Nvidia said Tuesday that Silicon Valley Power has sent more than 200 demand signals to an AI factory in Santa Clara, and that Emerald AI's software cut load from four megawatts to three while high-priority jobs kept running. Nvidia said Lambda's first validation found a fixed power budget can support 24 percent more token throughput when managed this way, and that Ian Buck made factory efficiency the center of his AI Infra Summit talk.
-- A factory that can shed a megawatt when the utility asks is making room for more clusters on the same grid, not sending the work home.
2026-09-15 / Nvidia / energy / central / #nvidia-factory-cuts-power-on-utility-signal
Periodic Labs said Tuesday it trained Periodic Neon, a trillion-parameter model, from the open-weight Kimi K2.6 using data from its Menlo Park benches, and that the last training run used 1,300 Nvidia H200 chips. Periodic said Neon reached 55.3 percent on FrontierXRD, a 20-fold gain from Kimi's 2.7 percent, and that it beat GPT-6 Astra and Claude Fable 5.1 at reading X-ray diffraction results at lower cost.
-- A lab that owns its experiments can beat a rented flagship on one task without 100,000 chips; that is still a lab, not a laptop.
2026-09-15 / Periodic Labs / efficiency / local / #periodic-labs-trains-neon-on-1-300-chips
Axis
local [#####################---------] central
Score: 69
clusters still winning
up 1
Google put live voice on its API. Salesforce kept a Nemotron model inside its own walls. Nvidia cut a factory from four megawatts to three on a utility signal. A Menlo Park lab trained on 1,300 chips. The cluster still holds the training run and the power.
Forces
| Force | Meter | Score | Pull | Note | ||
| CHIP / CAPEX |
|
86 | central | Google put live voice agents on its API. Salesforce and Nvidia put a CRM model on Salesforce's own machines. | ||
| ENERGY |
|
91 | central | Nvidia said a Santa Clara factory dropped from four megawatts to three on a utility signal so more clusters can fit the same grid. | ||
| EFFICIENCY |
|
48 | local | Periodic Labs said 1,300 H200 chips and its own lab data beat GPT-6 Astra at reading X-ray results. |
Line charts (0=local 100=central 2026-08-31 to 2026-09-16)
Market overlay (2026-08-18 to 2026-09-16)
Latest: BTC $75,941 / hashrate ~961.7 EH/s / token markers avg $5.3033/M / axis 69. These lines are different units. We compare timing, not dollars-to-hashrate.
Not shown: GPU cloud $/hour, hub electricity, miner hashprice, or lab cost per token -- none of those are a clean public feed.
Power and list prices
BTC $75,941 / hashrate ~962 EH/s. List-price gap (not a subsidy %): 44.4x list gap (blended) GPT-5.6 Luna -> Claude Fable 5 ($0.45/M to $20.00/M). (mempool.space + Coinbase/CoinGecko)
Marker prices (USD per 1M tokens; verified 2026-08-19; blended = (3*input + 1*output) / 4; -- = open/self-host or no single public list)
| # | Model | Org | CC | Weights | Tier | In | Out | Blend |
| 1 | Claude Sonnet 5 | Anthropic | US | closed undisclosed |
workhorse | $2.00 | $10.00 | $4.00 |
| 2 | GPT-5.5 | OpenAI | US | closed undisclosed |
flagship | $5.00 | $30.00 | $11.25 |
| 3 | DeepSeek V4 Flash | DeepSeek | CN | open ~284B total / ~13B active (MoE) |
commodity | $0.44 | $1.32 | $0.66 |
If useful, support the broadcast.
bc1qfs3yw5qlq8sxs50crzh3g84ug27gvww9npu85n
Subscribe
$ echo you@email.com | underclassic --subscribe
Previous issue | This week | Next issue | archive
One email per issue. No tracking pixels, no filler.
home | rss | page 2 | page 3 | archive | models | subscribe | data | board | mcp | llms.txt