AI infrastructure is consolidating while consumer AI services are learning that demand does not automatically produce healthy economics. Marvell has agreed to buy Lumentum for $115 billion to combine electrical and optical data center technologies. Anthropic is opening part of its IPO to Japanese investors, while OpenAI and xAI are tightening the way paid usage is metered. At the same time, startups are modifying Chinese open models for Western buyers and investigators are showing how deceptive apps can bypass platform safety reviews.
Marvell makes a $115 billion bet on optical AI infrastructure
Marvell agreed to acquire Lumentum in an all-stock transaction with an implied enterprise value of $115 billion. Lumentum shareholders would receive 4.5 Marvell shares for each share they own and hold about 33 percent of the combined company. The companies expect the deal to close in the first half of 2028, subject to shareholder and regulatory approvals.
The strategic logic is bandwidth. As AI clusters grow, copper connections face distance, power, and heat limits, pushing more data traffic onto optical links. Marvell brings switching, custom accelerator, and interconnect technology, while Lumentum adds lasers, optical transceivers, and specialized manufacturing. The companies claim a $550 billion addressable market and more than $1 billion in annual cost synergies. Those are projections, but the price shows how valuable connectivity has become beside the accelerators themselves.
Anthropic reserves up to $3.5 billion of IPO shares for Japan
Anthropic plans to offer Japanese investors up to 552.9 billion yen of shares, about $3.5 billion, as part of its US listing. A filing said the proceeds would support working capital, operating expenses, and capital spending. It did not state the full IPO size or valuation, although Reuters reported that some investors have discussed a valuation near $1.5 trillion.
The Japan allocation broadens direct access to an offering that is expected to be one of the largest technology listings on record. It also reinforces the capital intensity of frontier AI. Anthropic is not only selling investors a fast-growing software business. It is asking them to finance large and continuing commitments to models, compute, and distribution.
OpenAI combines ChatGPT and Codex usage allowances
OpenAI will place Plus, Pro, and Business users on a shared ChatGPT and Codex allowance in December. Usage will be measured in tokens, with consumption varying by model and task complexity. Instant chat will remain outside the shared pool. Subscribers who reach the limit can wait for a reset, upgrade, or buy credits for eligible features.
A unified meter simplifies billing as agentic Work features move into ChatGPT, but it also makes the cost of long coding and research tasks more visible. A single complex job can consume far more compute than a short conversation. The change gives OpenAI a common mechanism for balancing interactive chat, software agents, and premium workloads without promising the same volume for every kind of task.
Grok Bot exposes the tension between adoption and serving cost
Elon Musk said Grok Bot was too expensive to serve relative to its revenue after paid users complained that weekly usage limits were disappearing faster than expected. xAI has not disclosed cost, revenue, or quota figures, and it has not said whether it will change pricing or capacity.
The admission is useful because subscription growth can hide weak unit economics. Agent workloads may run longer, call more tools, and process more context than ordinary chat. Providers can respond through better models and infrastructure, but near-term levers are usually stricter limits, higher prices, or paid credits. OpenAI and xAI are approaching the same constraint with different billing structures.
Startups edit Chinese open models for Western deployment
Hirundo and Lazarus AI are modifying Chinese open-weight models to reduce censorship, security weaknesses, and behaviors that Western customers may reject. Hirundo says it edits model weights directly. Lazarus used prompts based on leaked censorship guidance and reference answers from Gemini Flash to realign Alibaba Qwen models, which it says are deployed at about a dozen US companies.
The work turns alignment into a supply-chain decision. Open weights let buyers inspect and change a model, but they also shift responsibility for evaluating hidden behaviors onto the deployer. Researchers caution that there is no universal Western alignment standard and that quick uncensoring methods can leave restrictions or misinformation intact. Enterprises need tests tied to their jurisdiction, users, and risk profile rather than a broad claim that a model has been made safe.
A deepfake network defeats app store review with bait and switch
A WIRED investigation linked more than 400 apps and domains to hidden abusive deepfake tools. Some apps appeared to offer pet portraits, language tutoring, or fitness coaching during review, then loaded explicit image and video generators under selected conditions. Researchers connected the network to Chinese-language developers and ads on Meta platforms. Google had removed most identified apps, while Apple removed 19 after being contacted.
The scheme exploited a gap between approving an app and monitoring what remote web content it later loads. Static review can miss behavior that changes by referral source, location, or server response. Platforms need continuous behavioral checks, shared infrastructure signals, and faster coordination across ads, payments, app stores, and hosting. Otherwise, generative abuse businesses can rotate disposable front ends faster than each platform investigates them.
What matters today
AI economics are becoming more explicit. Infrastructure companies are paying extraordinary prices for the components that move data, model labs are turning to public markets, and consumer services are metering expensive tasks more closely. At the same time, open-model customization and app-based abuse show that control does not end at model release. Sustainable deployment depends on connectivity, pricing, continuous testing, and enforcement that follows systems after launch.