Custom chip designer Marvell has signed a deal to acquire optical chip and networking equipment maker Lumentum in an all-stock transaction with an enterprise value of 115 billion dollars, including debt, and an equity value of about 80 billion dollars. Lumentum shareholders will receive 4.5 shares of Marvell common stock for each share they hold. TechCrunch, which reported the agreement on November 2, notes that the scale puts it among the largest acquisitions in technology history, ahead of Microsoft's 68.7 billion dollar purchase of Activision Blizzard and Dell's 67 billion dollar purchase of EMC.
The logic is about what actually limits a large AI cluster. Training and inference at frontier scale are no longer constrained only by how many accelerators a company can buy. They are also constrained by how fast data can move between those accelerators, between racks, and between buildings, and by how much power that movement consumes. Lumentum makes the lasers and optical components that carry data as light, while Marvell designs custom silicon and networking products. Marvell says the combination will make it a leader in AI connectivity, with products ranging from chips to data centers. That is a claim about owning more of the plumbing, not about releasing a better model.
One detail shows how central this layer has become. Nvidia is a major shareholder in both companies. In March, as part of a multi-year agreement to purchase Lumentum's advanced laser components, Nvidia invested 2 billion dollars to support Lumentum's research and development and fab expansion. Later that same month, Marvell announced that Nvidia had also invested 2 billion dollars in its optical networking and chips portfolio. When the dominant accelerator maker puts money into both sides of a deal, builders should read it as a signal that interconnect supply is strategic, not a commodity to be sourced at the last minute.
The deal is proposed, not completed. Marvell expects it to close in the first half of calendar year 2028, and only if shareholders of both companies and regulators agree. That long timeline matters. Anyone designing infrastructure today cannot plan around a combined Marvell and Lumentum product line that does not yet exist, and should keep second sources, contractual flexibility, and realistic delivery dates in view. Consolidation can improve integration, but it can also narrow choice in a market where lead times already shape what can be deployed.
A smaller release the same day pointed at the other end of the stack. Anthropic put message threads into public beta on its Messages API on November 2, allowing an application to send a full conversation once and then send only new messages with the previous response ID. Anthropic reports that median time to first token fell by about 35 percent on requests carrying 500,000 to 1 million tokens in its Claude Code rollout. The threads expire five minutes after the last request, token usage and pricing do not change, and developers must still keep their own copy of the conversation. Together, the two stories make the same practical point: the next gains for people building with AI will often come from moving less data, more efficiently, whether that data travels as light between chips or as a smaller request between an agent and an API.