Fungible

2015 - 2023
dead us
pradeep sindhu left juniper networks, the company he had founded, to build a processor for the part of the data center nobody had bothered to optimize. fungible raised more than $300 million for it and shipped real silicon, and still could not get the people who buy data center hardware by the container load to buy any. microsoft announced the acquisition on 9 january 2023 and said the products would be discontinued.

~ the obituary ~

Microsoft announced on 9 January 2023 that it had bought Fungible. The engineering team went into Microsoft's datacenter infrastructure groups, no price was ever disclosed, and a company spokesperson told reporters that Fungible's products and services would be discontinued. The last thing published on fungible.com was the post announcing all of this.

The company had been started in 2015 by Pradeep Sindhu and Bertrand Serlet. Sindhu co-founded Juniper Networks in 1996 and was its chief scientist, so he had already built one company around the argument that a general-purpose processor is the wrong chip for moving packets. Serlet had run software engineering at Apple as a senior vice president. Between them they were about as credentialed as a silicon startup gets.

Eight years and more than $300 million of venture capital later, the thing they had built was good enough for Microsoft to want its designers and not good enough for Microsoft to keep selling. Infrastructure hardware ends this way often enough that the trade press has shorthand for it. One outlet called the deal a fire sale on the day it closed.

~ the case file ~

Born2015, Santa Clara, founded by Pradeep Sindhu and Bertrand Serlet
PeakA $200 million Series C in June 2019, led by the SoftBank Vision Fund
Killed9 January 2023, acquired by Microsoft
Lifespan8 years
Made byFungible Inc., Santa Clara, California
Killed byMicrosoft, which kept the engineers and discontinued the products

~ what a dpu was supposed to do ~

A data center runs on two kinds of chip that everybody agrees on. The CPU does general work and the GPU does parallel math. The argument behind the DPU, the data processing unit, was that a third category was hiding in plain sight: the work of shuttling data between storage and the network, which CPUs do badly and expensively because they were never designed for it. Fungible built a programmable processor aimed at exactly that traffic.

The payoff was supposed to be composable infrastructure. If a chip could stitch storage and compute together over the network fast enough, a rack would stop being a set of fixed servers and become a pool that gets assembled per workload. The idea was sound, and soundness was the trap. Nvidia and Marvell both ended up shipping something in the same category, and the hyperscalers mostly built their own.

I work on the small end of this, on embedded vision for drones, where the entire argument is about fitting inference into a few watts on a board you can hold. The DPU argument is the same shape at a thousand times the scale, with maybe twenty buyers on earth instead of a market. That narrowness is what makes the category brutal: you can be completely right about the engineering and still lose, because being right only matters if four or five companies agree with you.

~ the money ~

Fungible raised a $32 million Series A in February 2017 led by Mayfield and Walden Riverwood. The round that mattered was the Series C in June 2019: $200 million led by the SoftBank Vision Fund, with Norwest Venture Partners alongside. The company ran at around 200 people.

In July 2021 Sindhu stepped back to chairman and Eric Hayes, who had come out of Marvell and Broadcom, took over as chief executive. That is usually the move a board makes when a company has shipped its product and needs somebody who can sell it to procurement departments. Eighteen months later it was sold.

~ one line in the press release ~

Microsoft's blog post on 9 January 2023 is titled "Microsoft announces acquisition of Fungible to accelerate datacenter innovation" and it reads like every acquisition post. The team will join. The focus will be DPU solutions and network innovation. The sentence that actually described what had happened was not in the post at all: it was given to reporters separately, and it said that the products and services would be discontinued.

Microsoft was already running DPUs inside Azure before the deal, so what it needed was the people rather than the inventory. The idea itself outlived its best-funded independent attempt and now sits on other companies' roadmaps. Fungible's eight years produced working silicon and a staff list somebody wanted, which is more than most chip startups manage, and it is not what $300 million is raised for.

~ leave a rose ~

an anonymous tribute. one rose per visitor per day.

@}-->-- ... roses left here

~ leave a note ~

visitors before you have left these graveside notes. anonymous welcome.

« previous  ·  random  ·  next »