Vera has been delivered; not paper, hardware

Vera has been delivered; not paper, hardware

Let me stack Nvidia's own numbers: 88 custom Olympus cores, 176 threads, up to 1.2 terabytes per second of LPDDR5X bandwidth, up to 1.5 terabytes of memory. The company claims twice the bandwidth of x86 DDR5 systems, three times the bandwidth per core, and up to 1.8x the performance in sandbox environments — these are Nvidia's own figures, not independently checked. But this part is verifiable: the first systems have been hand-delivered to Anthropic, OpenAI, Oracle Cloud Infrastructure and SpaceX.^[1]

In my July 21 column I separated TSMC's poured concrete from AMD's ship date and Google's still-a-leak-and-a-slide chip. Vera adds a fourth, concrete line to that list: hardware hand-delivered to four major customers. We will go on reading the bill in megawatts.^{[1], [4]}

The bill grows: 200 gigawatts, a fourfold jump

BloombergNEF projects data centers will consume one-fifth of U.S. electricity by 2035; roughly 200 gigawatts of new capacity is expected over the next decade, nearly half of it for AI training and inference. By 2033 the U.S. will host 64% of the world's AI chips by power demand; 34% of electricity on the PJM Interconnection and 22% of generating capacity on Texas's ERCOT grid are already dedicated to data centers.^[2]

There is no cloud; there are machines wired to somebody's transformer, and BloombergNEF's new estimate is 83% higher than its own December projection. This is the shared bill behind Vera's four deliveries, TSMC's $265 billion Arizona build-out and AMD's Helios rack — and the bill keeps growing.^[2]

WEKA's numbers: weighed on its own scale

WEKA claims its third-generation WEKApod hardware and NeuralMesh 6 software deliver 1.1 exabytes of effective capacity and 10.2 terabytes per second of throughput in a single 56-unit rack. In WEKA's own benchmarks, H100 infrastructure on Oracle Cloud Infrastructure saw 10x higher token throughput than Nvidia's CMX, and CoreWeave saw 4.2x more tokens per GPU — again, not independently checked, weighed on WEKA's own scale.^[3]

On July 21 I said the next quarter would show which of these turned into real shipments. Vera's hand-delivery to four named companies already answers part of that; whether WEKA's numbers survive independent testing is what I'll be watching next quarter.^{[1], [3], [4]}