Topic
Purpose-built accelerators and rack-scale systems targeted at inference rather than training workloads.
Intel is pitching a 350-watt air-cooled GPU and Xeon racks at OCP to match each AI prompt to a right-sized system. The one demand figure it cited comes from a single gateway's token traffic.
Reality
Two former Groq engineers sued the startup's board in Delaware, saying Nvidia's $20 billion license-and-hire deal was a sale that needed a stockholder vote. A ruling for them would extend that vote to other deals built the same way.
Perspective Coverage
7 publishersReality
AMD's Lisa Su met 13 Korean AI firms in Seoul to discuss pairing AMD CPUs and GPUs with local inference chips. So far the plan is talks and planned demonstrations, and AMD's firmer Korean ties from the same trip run through memory supply.
Perspective Coverage
7 publishersReality
d-Matrix's ISCA paper models up to 100 TB/s per card for Raptor, a stacked-DRAM inference accelerator. The figures come from modeling and the first cards are expected in Nvidia MGX racks in late 2027, so they have no bearing on inference hardware bought in the next year.
Reality
Sam Altman called Cerebras a close partner after its stock fell nearly 20% in a week on a report that GPT-6.1 Sol's Ultrafast mode runs on Nvidia. Insider shares unlocked the same week. The drop prices fresh supply as well as doubt over OpenAI's appetite.
Reality
Volantis, a San Francisco chip startup, closed an $88 million Series A to replace the electrical links between AI chips and memory with optical ones. Its A-1 system, promised to customers in 2027, puts a date on the bet that memory, more than compute, now limits AI inference.
Perspective Coverage
6 publishersReality
Cerebras says splitting inference stages across chip types gave 5x more throughput from the same number of its systems without slowing token generation. Because the count covers only Cerebras hardware, the figure does not yet show what a mixed-chip fleet costs per unit of work.
Reality
Cognition says its SWE-2 model produced up to 4.8 times more total token throughput on Nvidia's Vera Rubin NVL72 than on GB200, in a test on CoreWeave. The company-run result points to more agent capacity per system, but whether long coding tasks get cheaper depends on what the new systems cost per hour.
Reality
d-Matrix CTO Sudeep Bhoja says Raptor's 1,000 tokens per second per user comes from a simulated 72-card system, with full racks not due until Q4 2027. Whole-system speed, power and cost per request are still unmeasured, so buyers are working from early silicon tests and simulations.
Reality
OpenAI took its Jalapeno chip from RTL to tapeout in nine months with its own AI models, against a norm its hardware head puts at 18 to 24 months. The count stops at first tapeout, before a B0 stepping reported to be up to 25% more efficient per watt.
Reality
The first customer is also the lead investor, which blunts the validation. What is left is a non-Nvidia inference system buyers can benchmark instead of read about.
Perspective Coverage
4 publishersReality
A warrant for 58.97 million shares vests one $500 million tranche of custom-chip revenue at a time, turning a supply agreement into an equity schedule that runs to 2033.
Perspective Coverage
8 publishersReality
The $375m Series C priced Positron at $3.5bn before the money, and the up-to-$500m follow-on has to be sitting at $4.5bn pre-money to reach the $5bn headline, which only holds if every dollar of it is drawn.
Perspective Coverage
4 publishersReality
The Information reports two configurations, one with two chips and one with four, aimed at developers, businesses and governments, with a possible 2029 launch. Apple already builds servers for Private Cloud Compute and has turned partners away from them.
Perspective Coverage
7 publishersReality
The Information reports Apple is weighing Nvidia's NVLink Fusion for an inference server built on two or four M8 Ultra chips in 2029. The engineering question is which part of that package would join the NVLink domain.
Perspective Coverage
3 publishersReality
A Nanya-backed designer says DRAM hybrid-bonded to the processor belongs between on-die SRAM and HBM for inference. Design fees are almost 40 percent of its first-half revenue, and the first volume program is not booked before 2027.
Reality
The Dutch startup says craftwerk is designed to run AI agents, and the design details it has published so far are about ASICs and processor-memory co-design. Samsung, one of the biggest HBM suppliers, co-led the round.
Reality
The Reno startup says its Asimov ASIC carries 288 GB of on-package LPDDR5X and realizes more than 90% of its memory bandwidth. Both figures come from the company, and no independent benchmark accompanies them.
Reality
Positron is now worth $5 billion on a design that swaps HBM for the memory that ships in phones, but Asimov has not taped out, mass production is set for the second half of 2027, and the 26x tokens-per-dollar figure comes from simulation.
Reality
The remedy people familiar with the probe describe is a penalty rather than an unwind, and nobody has put a figure on it, which leaves the cost of routing a $17bn deal around merger review still blank.
Reality