Monday, September 14, 2026

Monday’s September 14 trading snapshot put Nvidia down 3%, AMD down 5.7%, and SK Hynix down 6.3%. The move followed Amodei’s September 12 call for a slower pace of model development and support from Altman and Musk. These are reported intraday moves, not closing returns.

New AgentX measurements published September 14 put Vera Rubin NVL72 at 1.4–3 times the throughput per total cost of ownership of GB300 running TRTLLM at 60–100 tokens per second. The report's 67× result applies at 170 tokens per second against GB300 Dynamo TRTLLM under owning-cost assumptions. The measurements use early pre-release software.

New September 14 conference checks point to another couple of years of laser shortages at Lumentum and 26–28-week optical lead times at MaxLinear. The note quotes 1.6T DSP pricing of $80–100 versus $35–50 for 800G. This is new channel reporting, with the product transition still ahead.

TrendForce's first Weekly Radar, published September 14, identifies severe shortages in three of its six tracked component categories. ABF substrate lead times are 48–56 weeks against a 12-week balanced-market baseline. GPUs are the only tracked category at a balanced lead time.

The September 14 reconstruction sustains 12 robots per B300 within a 500-millisecond deadline, versus four on an RTX PRO 6000. It measures the cost of serving a RoboTTT-shaped workload, not robot task success. The accompanying cost model favors shared inference when server utilization stays high.

• Reuters reports a global selloff in AI-linked stocks on September 14 after leading AI executives called for slower capability development. Its trading snapshot put Nvidia down 3%, AMD down 5.7%, and SK Hynix down 6.3%.?
• Reuters reports that Dario Amodei called on September 12 for embedded independent evaluators, coordination among frontier labs, and international cooperation. Sam Altman and Elon Musk expressed agreement. Altman supported evaluators with employee-like access.?

• SemiAnalysis reports 1.4–3× Rubin throughput per dollar versus GB300 running TRTLLM at 60–100 tokens per second in new September 14 AgentX results using early pre-release software.?
• SemiAnalysis reports roughly 67× throughput per total cost of ownership at 170 tokens per second against GB300 Dynamo TRTLLM under its owning-cost assumptions.?

• FUNDA reports in its September 14 conference note that Lumentum expects laser shortages to last another couple of years, with new orders at already higher prices or above.?
• FUNDA reports MaxLinear optical lead times of 26–28 weeks. Its conference checks put 1.6T DSP prices at $80–100 versus $35–50 for 800G, with the transimpedance amplifier taking 1.6T content above $100.?

• TrendForce reports severe shortages in three of six tracked component categories in its first Weekly Radar, published September 14.?
• TrendForce reports 48–56-week ABF substrate lead times versus a 12-week balanced-market baseline, with GPUs the only category at a balanced lead time.?

• SemiAnalysis reports that one B300 GPU sustains inference for 12 robots within a 500-millisecond deadline, versus four on an RTX PRO 6000, in its September 14 timing-representative RoboTTT reconstruction. The test measures serving cost, not robot task success.?
• SemiAnalysis estimates B300 offload at about 46% of Jetson Thor cost per unit of compute in its base case, assuming roughly 90% server utilization versus 40% on-device utilization. The result depends on pooling robots and maintaining high server use.?

• SemiAnalysis finds in its September 13 roofline analysis that taller HBM stacks add less Kimi K3 throughput than system cost under the modeled assumptions. It estimates 8-high and 12-high system-cost premiums of 12.1% and 26.3% over a 4-high HBM baseline. This is a model of a future system, not a delivered hardware benchmark.?
• SemiAnalysis reports a separate 16-GB300 experiment in which reduced HBM preserved similar throughput over much of the curve but lost almost 30% above concurrency 70. The result shows the limit of using DRAM cache offload to reduce HBM needs.?

• Electron Economics argues in its September 14 analysis that long power-fleet contracts expose buyers to a mismatch: much of the planned engine-supply relief clusters in 2028–2030, but contract terms can extend to 18 years. The new contribution is this duration-risk comparison. The capacity plans and contracts predate the report.?
• Electron Economics argues that expanding engine supply would compress the premium paid for early delivery, with service economics more durable. It explicitly allows for delays from castings, forgings, test facilities, and staffing constraints.?