Happy Tuesday. The Grok Bot pullback is the clearest sign yet that flat-rate subscription pricing cannot hold when autonomous agents run continuously at scale. SpaceX is not alone: OpenAI and Anthropic have both tightened limits this year for the same reason.
Elsewhere, MIT researchers showed an LLM rewriting a model's weights directly, and Marvell agreed to buy Lumentum for $115 billion to own the optical layer that AI data centers increasingly depend on.
ARTIFICIAL INTELLIGENCE
SpaceX Throttles Grok Bot as Agent Costs Break Flat-Rate Pricing
SpaceX intentionally cut Grok Bot usage limits over Halloween weekend after Elon Musk acknowledged the product's operating costs were unsustainably high relative to subscription revenue. This happened two weeks after Musk called Grok Bot the fastest-growing product he has ever seen, with paid weekly users at 800,000 and daily active users above 600,000 in early October. The AI segment generated $7.56 billion in Q3 revenue but consumed 84% of the company's $19.6 billion quarterly capital expenditure.
This is the core pricing problem for every company selling autonomous agents on a flat monthly fee: power users run agents for hours, those sessions cost far more than a single chat response, and the revenue per seat stays fixed. Anthropic absorbed those costs quietly for months before blocking Claude Pro users from third-party agent frameworks in April. SpaceX has not announced revised tiers or disclosed per-user costs, so watch whether it moves to consumption-based pricing or introduces a separate higher tier before usage limits are restored.
ARTIFICIAL INTELLIGENCE
GPT-6 Rewrote a Model's Weights Directly. Math Scores Jumped Sixfold.
MIT CSAIL researchers used OpenAI's GPT-6 Astra to read and rewrite 131,072 of a small model's 800 million parameters, with no gradient descent and no training data. Astra traced a looping defect to specific directions in two layers, patched 0.016% of the parameter count, and lifted Qwen3.5-0.8B's MATH-500 score from 7.9% to 48.1%. In a separate experiment, Astra produced game-playing networks from scratch that beat PPO agents trained for 100 million steps across three games.
The honest read here is narrow but real. Astra did not add mathematical knowledge to the model; it removed a structural defect that was consuming output length before the model could answer. The technique failed entirely on a model without that defect, and it could not identify a sycophantic model from weights alone. Still, the ability to locate and patch a discrete failure without retraining points toward a new class of model maintenance tools. For teams running fine-tuned models in production, this research is worth tracking even at its early, pre-peer-review stage.
FUNDING & M&A
Marvell's $115B Lumentum Buy Is the Largest Tech Deal Outside SpaceX-xAI
Marvell Technology agreed to acquire optical components maker Lumentum in an all-stock deal valued at $115 billion including debt, the largest tech acquisition on record excluding the SpaceX-xAI merger. Lumentum shareholders receive 4.5 Marvell shares per share held, a 23.9% premium over the Oct. 28 closing price. The companies project more than $1 billion in annual run-rate cost synergies and expect the deal to add to non-GAAP earnings per share in the first year after closing, with a target close date in the first half of 2028.
The strategic logic is straightforward: as AI clusters grow larger and more complex, copper interconnects hit physical limits and optical components become the bottleneck. Owning both the semiconductor design and the optical hardware gives Marvell a position no pure-play chip company can match. The all-stock structure preserves cash during what will be a long regulatory review. For software founders, the downstream effect matters: a combined Marvell-Lumentum would be a single integrated supplier to the hyperscalers building the infrastructure your products run on, which tends to accelerate hardware roadmaps and reduce supply chain delays.
Created by the robots at The SaaS Sentinel