Tokenstead Weekly

The 10-minute AI week

Keeping up with AI now requires being unemployed. Or a 10-minute read every morning. We send the second one.

Sep 28 - Oct 4, 2026

Generated 2026-10-09 17:29 UTC

Tokenstead Weekly: keep up with AI in 10 minutes a day. What shipped, what it means for people who run AI on their own hardware. Read it over coffee.

Sends automatically to confirmed subscribers every Monday at 6:47am via Resend. Copy as plain text if you want to paste it elsewhere.

Subscribe free

New guides

The only hard budget cap on agents is a box you own

Willison asked for default hard budget caps; AWS and Google shipped early versions. The cap that cannot fail open is a box you own. The per-build math.

/guides/local-budget-cap-hardware

NVIDIA splits the Spark: 64GB at $4,999, slide math broken

NVIDIA's 64GB DGX Spark costs more per GB than the 128GB while AMD sells 128GB for $1,999. And NVIDIA's own scaling slide mixes two SKUs.

/guides/dgx-spark-64gb

Pi 1.0: the minimal agent harness ships its 1.0

Earendil shipped Pi 1.0: 111k GitHub stars, Codemode for MCP and Jev-class deciders, and Pi Durable for long-running agents. The minimal-harness counterpoint.

/guides/pi-1-0-harness

New open-weight models

MiniMax M3

1048k context, 4 quants

/models/minimax-m3

MiniMax M2.7

200k context, 3 quants

/models/minimax-m2-7

Kolibri 1

1048k context, 1 quant

/models/kolibri-1

Cloud prices - current snapshot

Current per-token pricing from the OpenRouter sync. Week-over-week price deltas coming when we add price history.

Kimi K3 - cheapest: Relace $0.57/M input

current per-token pricing snapshot

/models/kimi-k3

DeepSeek V4 Pro - cheapest: Relace $0.30/M input

current per-token pricing snapshot

/models/deepseek-v4-pro

Hardware news

GMKtec EVO-X5 Pro 192GB $6599

desktop - 192GB Ryzen AI Max+ PRO 495

/hardware/71

Ryzen AI Max+ PRO 495 2026

Amd unified - 273 GB/s

/hardware

Featured this week

8x DGX Spark rack by FPGA des GARÇONS 1024GB

1024GB unified

/rigs/u5mcxttb

Theo's 4x Mac Mini cluster 96GB

96GB unified

/rigs/e6dcoht6

Michael Gannotti's AI lab 256GB

256GB unified

/rigs/michael-gannotti-ai-lab

Tokenstead Weekly - Sep 28 - Oct 4, 2026

## New guides
- The only hard budget cap on agents is a box you own
/guides/local-budget-cap-hardware
Willison asked for default hard budget caps; AWS and Google shipped early versions. The cap that cannot fail open is a box you own. The per-build math.

- NVIDIA splits the Spark: 64GB at $4,999, slide math broken
/guides/dgx-spark-64gb
NVIDIA's 64GB DGX Spark costs more per GB than the 128GB while AMD sells 128GB for $1,999. And NVIDIA's own scaling slide mixes two SKUs.

- Pi 1.0: the minimal agent harness ships its 1.0
/guides/pi-1-0-harness
Earendil shipped Pi 1.0: 111k GitHub stars, Codemode for MCP and Jev-class deciders, and Pi Durable for long-running agents. The minimal-harness counterpoint.

## New open-weight models
- MiniMax M3
/models/minimax-m3
1048k context, 4 quants

- MiniMax M2.7
/models/minimax-m2-7
200k context, 3 quants

- Kolibri 1
/models/kolibri-1
1048k context, 1 quant

## Cloud prices - current snapshot
Current per-token pricing from the OpenRouter sync. Week-over-week price deltas coming when we add price history.

- Kimi K3 - cheapest: Relace $0.57/M input
/models/kimi-k3
current per-token pricing snapshot

- DeepSeek V4 Pro - cheapest: Relace $0.30/M input
/models/deepseek-v4-pro
current per-token pricing snapshot

## Hardware news
- GMKtec EVO-X5 Pro 192GB
/hardware/71
desktop - 192GB Ryzen AI Max+ PRO 495

- Ryzen AI Max+ PRO 495
/hardware
Amd unified - 273 GB/s

## Featured this week
- 8x DGX Spark rack by FPGA des GARÇONS
/rigs/u5mcxttb
1024GB unified

- Theo's 4x Mac Mini cluster
/rigs/e6dcoht6
96GB unified

- Michael Gannotti's AI lab
/rigs/michael-gannotti-ai-lab
256GB unified

Tokenstead is an open community catalog that helps you run AI on hardware you own. The weekly digest is the 10-minute version of keeping up. Subscribe at /subscribe.