|
|
|
|
|
|
Nvidia fell 2.2%, AMD fell 1% - and the market read the trade backwards
|
|
JULY 20TH, 2026 · BY JONATHAN HILDEBRANDT
|
|
Nvidia fell 2.2% and AMD 1% on July 20 as traders sold chip stocks, betting Kimi K3's efficiency would dent GPU demand, a replay of the January 2025 DeepSeek panic. Bloomberg's July 20 analysis says that read is backward. Kimi K3 is a 2.8-trillion-parameter sparse MoE with a 1M-token context, and its agentic loads are memory-bandwidth-bound. That pulls HBM demand toward SK Hynix and Micron, it does not pull volume away from GPUs.
Why it matters. The bottleneck shifted from compute to memory, which changes what you actually pay for on long agent loops. Our Kimi K3 review has the number the whole thesis rests on: cache-hit input at $0.30 per million tokens and a 42.0 SWE Marathon score against Fable 5's 35.0 (per MarkTechPost). See the pricing math.
|
|
See the pricing math and the exact workload where switching pays off →
|
|
|
Netflix paid $587M for Ben Affleck's AI post-production startup
JULY 20TH, 2026 · PONDERO NEWSDESK
The target is InterPositive, a 16-person team building relighting, VFX compositing, and color-grading tools, per an SEC filing Variety confirmed July 19. Read what it is not: this is the editing pipeline, not a generative-video play. Netflix already runs GenAI across 300-plus titles and just bought the finishing room, not the camera. Variety.
|
Apple sued OpenAI over trade secrets, and there is still no injunction
JULY 20TH, 2026 · PONDERO NEWSDESK
Apple's July suit alleges former VP Tang Yew Tan showed OpenAI engineers actual hardware prototypes, with 400-plus ex-Apple staff now on the roster. OpenAI says its device stays on track for a 2026 reveal and 2027 release. Nothing filed so far blocks the product, so treat the timeline as intact until a judge says otherwise. TechCrunch analysis.
|
DeepSeek retires its V4 aliases on July 24
JULY 20TH, 2026 · PONDERO NEWSDESK
Any call using deepseek-chat or deepseek-reasoner starts failing at 15:59 UTC on July 24; rename to deepseek-v4-pro or deepseek-v4-flash before then. Peak pricing lands the same day and roughly doubles cost during Beijing business hours (09:00 to 18:00 CST, weekdays), so schedule heavy batch jobs off-peak. Migration guide.
|
Cursor now shows its plan before it touches your repo
JULY 20TH, 2026 · PONDERO NEWSDESK
Triggered from Slack, Cursor posts a plan first so the team can redirect before any code changes, streams live status, and runs named multi-repo environments (frontend, backend, and a shared lib as one target). New Cloud Agent hooks land too: beforeSubmitPrompt, afterAgentResponse, stop. Fewer surprise pull requests is the real win here. Changelog. Try Cursor.
|
|
•
|
Cloudways. Managed cloud hosting for the servers this week's news implies you might need: a box to test self-hosted Kimi K3 inference, or a home for an n8n or Make agent. Spin one up in about five minutes. Try Cloudways →
|
|
•
|
Buttondown. The platform sending you this email. Clean API, no AMP, and pricing that does not punish list growth. If you want off Beehiiv or Substack, this is the switch we made. Try Buttondown →
|
|
|
|
|
Jonathan Hildebrandt
Co-founder and primary operator of Pondero. Writes the Pondero Brief.
|
|
|
Affiliate disclosure
·
Unsubscribe
·
Manage preferences
Pondero earns commissions on some links. This does not affect our editorial picks.
|
|
