AIGoogle Paid Half a Billion for DeepMind. It Bought Time.
Google has acquired DeepMind, a London research company, for a figure reported at around four hundred million pounds.
DeepMind has no product. It has no revenue. It has, depending on which account you read, somewhere between fifty and seventy-five employees.
Work out the price per head and then let’s talk about what that number means.
This Is Not an Acquisition
Companies get bought for their customers, their revenue, their technology, or their people. This is entirely the fourth kind, and the fourth kind behaves differently from the other three.
When you buy revenue, you’re buying something you can model. Discount the cash flows, apply a multiple, argue about the growth rate. When you buy a research team, you’re buying the option to be at the front of a field, and the price is set by how many other bidders want the same option.
The reporting suggests Facebook was also interested. That’s the entire explanation for the number. There is no discounted cash flow here. There is a small number of groups in the world who can do this work, a smaller number available for purchase, and several enormous companies who have concluded they cannot afford to be without one.
“Nobody paid four hundred million pounds for fifty people’s output. They paid it for the certainty that a competitor wouldn’t have them.” — Sameer Gupta
What It Says About Where Capability Sits
Here’s the part I’d have a strategy conversation about.
The techniques in this field are, for the most part, published. The papers are public. The mathematics is not secret. The ImageNet result that started this wave was described in full detail and anyone can read it.
So why can’t a large company with money simply hire ordinarily good engineers and reproduce the work?
Because the gap between reading a paper and reproducing its results is enormous, and it is filled with judgement that nobody has written down. Which architecture to try. How to initialise. What to do when the loss stops falling. Which failures mean the idea is wrong and which mean you’ve made a mistake. That knowledge lives in a few hundred people’s heads and it is transmitted by working alongside them.
That’s why the price is what it is. The published papers are the visible part. The tacit knowledge of how to make them work is the asset, and it can only be acquired by acquiring people.
The Implication for Everybody Else
I want to draw the uncomfortable conclusion rather than dance around it.
If Google and Facebook are bidding four hundred million pounds for fifty researchers, then your company is not going to hire a frontier research team. Not at any price you would sanction, and not against that competition. That door is closed, and pretending otherwise wastes a hiring budget.
This is fine. It’s also clarifying, because it tells you what your actual strategy has to be.
- You are a consumer of this research, not a producer of it. Your advantage will come from applying published work to a problem you understand better than anyone else, not from advancing the state of the art.
- The lag is your friend. Techniques that required a research team in 2012 are becoming library calls now. Whatever DeepMind is doing today will be in an open source package in three or four years. Waiting is a legitimate strategy and it is dramatically cheaper.
- Your data is the part they can’t buy. Google can hire anyone. It cannot acquire your ten years of maintenance records, your claims history, or your customers’ behaviour in your particular market. That asset is genuinely yours and it’s the one worth investing in.
“You will not win the talent auction. You don’t need to. You need to own a dataset and a problem that the people who won it have no access to.” — Sameer Gupta
What I’d Watch
A few things I’d expect over the next several years, offered with the usual caveat that I’m reading this from the outside.
The concentration gets worse before it gets better. The handful of labs that can do this work are being absorbed into a handful of companies. That has consequences for who sets the research agenda and what gets published, and I’d expect the openness of the last few years to narrow somewhat.
Talent costs cascade downward. When the top of the market prices at these levels, the merely-very-good become expensive too. If you have people who understand this material, they are now worth substantially more than your compensation bands say, and somebody will tell them.
Applied beats frontier, commercially. The research is glamorous. The money, for most companies, will be in unglamorous application of well-understood techniques to specific business problems. That’s less exciting and considerably more achievable.
Final Thoughts
The interesting thing about this acquisition isn’t the price. It’s that the price was set with no reference to what the company produces, because it doesn’t produce anything yet.
That’s a market pricing pure capability, and pure capability in this field currently lives in a small number of people rather than in any institution or technology. Which means it’s mobile, which means it’s contested, which means it’s expensive.
For anyone running a normal business, the takeaway is not to compete. It’s to be very clear that your advantage lives in your data and your domain, and to build accordingly.