Habana Labs: $2 billion in the press release, $1.7 billion in the annual report
On 16 December 2019 Intel announced that it had acquired Habana Labs, an Israeli developer of programmable deep learning accelerators for the data centre, for "approximately $2 billion". Eight weeks later the company's annual report for fiscal 2019 named the same deal in one sentence and with a different number: approximately $1.7 billion.
Why it matters
A $300 million gap between a press release and the audited annual report of the same company for the same quarter is the cheapest possible test of how much an announcement figure weighs. Both documents are primary, both come from the buyer, and they differ not in interpretation but in the sum. The first is the one that circulates; the second is the one that passes audit.
The press release calls Gaudi a training processor then sampling with select hyperscale customers and Goya an inference processor already commercially available, and says systems built on Gaudi are expected to deliver up to four times the throughput of systems built with an equivalent number of GPUs. The annual report repeats both claims almost word for word. The release adds Intel's own forecasts: over $3.5 billion of AI-driven revenue for 2019 and an AI silicon market above $25 billion by 2024, of which more than $10 billion in the data centre. What this record does not assert. The fourfold advantage over GPUs is a manufacturer's estimate of its own product, and this record carries no independent measurement of it. The record does not assert that Intel cancelled the Nervana line at the same time: Nervana occurs twice in the 2019 annual report and both times as a current product. The string $2.0 billion occurs three times in that report and never about Habana — it is the 2009 debentures, non-US pension obligations and the fair value of vested restricted stock.