In February 2021, an electricity market presented Matt Wytock’s company with a peculiar invitation: prices were extraordinarily high, and its models were unprepared. During the Texas freeze, Gridmatic decided to sit out the first part of the week. A business built to forecast prices had encountered a situation in which restraint mattered more than another prediction.
The company’s account was unusually direct. Its models had not been built for sustained prices of $9,000. Sitting out was an option, and it took it. For a founder selling the usefulness of machine learning, acknowledging that boundary made an interesting calling card. Even a clever algorithm occasionally needs someone to close the laptop on its behalf.
Wytock’s career makes the episode worth examining. He worked on Google’s core Search and Ads systems, created the product known as Keywordless Ads or Dynamic Search Ads, earned a machine-learning doctorate, and founded Gridmatic in 2017. His current work places mathematical predictions inside an operating electricity business. A forecast can influence a battery’s next move and a customer’s power bill.
There is a tempting version of this biography in which a software engineer arrives in an old industry and teaches it to count. The more interesting story contains mentors, years of energy research, experienced commercial colleagues and a moment when the machinery declined to play. Electricity has a useful way of making intellectual confidence expensive.
A Google colleague’s inconvenient advice
Wytock’s route into graduate study began at Google. R. V. Guha, whose project he joined on his first day, encouraged him to pursue a doctorate. He later credited Guha with shaping his approach to research. At Carnegie Mellon, adviser Zico Kolter introduced him to energy problems; Stephen Boyd hosted him at Stanford. The move to Pittsburgh also involved his wife, Audra, who relocated from California to support his studies.
Those connections give the career a different texture from the familiar tale of a solitary founder discovering an overlooked market. There were people directing his attention toward further study, toward electricity, toward questions outside a conventional computer-science brief. Advice can be inconvenient when it asks for years rather than an afternoon. In Wytock’s case, it became part of the work.
The advertising product offers one way to understand his earlier engineering experience. Dynamic Search Ads automates parts of a process that would otherwise require advertisers to maintain keyword choices. The broad comparison with electricity is an editorial one: both involve decisions made repeatedly, with more information than a person can comfortably review by hand. The resemblance ends quickly when a battery runs out of charge.
His academic record shows that energy was already a serious subject well before Gridmatic opened for business. In 2013, Wytock and Kolter published research on energy forecasting at the International Conference on Machine Learning. They developed faster methods for learning relationships between inputs and outputs and applied them to real energy tasks. Electricity had entered his professional life through research, years before it became the company’s product.
The household hidden inside a number
Another 2013 paper, also with Kolter, examined energy disaggregation. A smart meter can report a household’s combined electricity use without telling you which uses contributed to it. Their research used contextual information to estimate components of that aggregate signal. The work covered thousands of homes across four years. It required extracting something useful from measurements that did not arrive neatly labeled.
That is a less glamorous assignment than asking a computer to paint a picture, but it brings the stakes of applied machine learning down to a recognizable scale. One number can conceal several activities. Better interpretation can make that number useful. Wytock was working on the relationship between available data and practical understanding before generative AI became a routine dinner-party topic.
In 2016, he coauthored a paper with Steven Diamond, Felix Heide and Boyd on optimization frameworks. The proposal expressed solvers as computation graphs, using systems such as TensorFlow, to support hardware including GPUs and distributed clusters. It introduced the open-source Python framework cvxflow. The subject was computational plumbing: how to let a mathematical model run efficiently on the machines available to it.
By 2019, he was also a coauthor of a chapter on dynamic energy management with Nicholas Moehle, Enzo Busseti and Boyd. Its framework used forecasts to choose current power flows, then accounted for uncertainty in those forecasts. Read together, these projects suggest a coherent technical progression: interpreting energy data, making computation practical, and turning a view of the future into a decision in the present. That progression is visible in the published work.
A battery with an appointment book
Gridmatic brings those kinds of decisions into electricity markets. Its business spans trading, battery operation, renewable supply and retail energy. The question for a battery is easy to state: when should it charge, when should it discharge, and what should it commit to doing before that moment arrives? Answering it requires forecasts and the discipline to act within physical and commercial limits.
In an April 2023 interview, Wytock described a model of the United States electricity grid down to individual market nodes. He also described a storage backtest, validated by DNV, that showed a potential average revenue uplift of 28% across ERCOT systems. That number belongs to a backtest. It should retain its label, rather than acquire a new identity as a promise about every battery.
His November 2023 essay explained why an operator’s job keeps changing. As batteries competed to provide a grid-balancing service called Regulation Down, the prices for that service declined. Owners needed to find other sources of revenue, including energy sales in real time. The battery could remain physically unchanged while the business around it shifted. Buying equipment did not buy a permanent strategy.
The same essay distinguished actual revenue, revenue achievable with perfect knowledge, and potential revenue under AI optimization. The distinction is useful because hindsight is a remarkably well-paid imaginary employee. Real operators must decide before prices reveal themselves. The interesting measure is what their system earns with information available at the time, while honoring the commitments it has already made.
A later Gridmatic analysis made a related point about benchmarks. In August 2025, its Endurance Park battery achieved a reported 115% capture rate against a real-time price-spread benchmark. Exceeding 100% did not mean defeating arithmetic: the benchmark used simplified assumptions about hourly prices and one daily cycle. Actual participation offered finer intervals and other opportunities. A percentage becomes more informative when someone explains its denominator.
Potential average storage revenue uplift in the ERCOT backtest Wytock described in 2023, validated by DNV. A historical simulation, not a guaranteed return.
Clean power needs a clock
In January 2023, Gridmatic launched its retail business for commercial and industrial customers, initially in Texas. Among the announced customers was EdgeConneX, seeking electricity matched to carbon-free supply hour by hour. Wytock’s company was moving closer to the customer’s meter. A forecasting business now had to turn its calculations into a product someone could procure and use.
Hourly matching makes the clock part of the clean-energy claim. A data center keeps consuming after solar output falls. Counting renewable generation over a longer period can miss that mismatch. Serving a customer through each hour calls for a suitable mix of supply, storage and operating decisions. For Wytock, this extends the task from predicting a market to assembling a service around its changing conditions.
The following month, Gridmatic announced an agreement to buy 15 megawatts from Avangrid’s Barton Chapel wind project in Jacksboro, Texas. The agreement covered 2023 and 2024 and supplied its retail customers. It was a concrete transaction behind the software discussion: a named wind farm, a defined quantity and a contract period. The forecast had acquired counterparties.
By September 30, 2026, Gridmatic reported more than 1,600 commercial and industrial retail customers signed over two years. Its retail arm also announced recognition for price competitiveness in Energy Research Consulting Group’s annual broker and consultant study. These are company milestones under Wytock’s leadership. They describe a business growing beyond research papers into the repetitive, consequential work of supplying customers.
The colleagues behind the calculations
Wytock’s public hiring announcements offer a view of the people he assembled around that business. In 2022, he welcomed Diamond, the author of the convex-optimization modeling language CVXPY, alongside Jamie Ray and Jarrett Hunt. Diamond had worked on portfolio construction at BlackRock; Ray brought experience forecasting prices with deep learning. These appointments connected mathematical tools with the practical problem of deciding what to do in a market.
The connection with Diamond also reaches back to their 2016 paper. Research collaborations and company hiring overlap in this story. That overlap is worth noticing without pretending a coauthor alone explains a business. Gridmatic also has leaders with backgrounds in energy storage, project finance and retail supply. The machine-learning founder’s organization has to contain knowledge that cannot be reduced to his own résumé.
The company’s published engineering principles describe engineers taking responsibility from concept through infrastructure and product design. They also emphasize reproducible backtests, risk scenarios and learning the energy domain. That is an account of how the organization wants to work, rather than a personality assessment of its founder. It nevertheless gives a practical setting for his decision to build an operating company around research.
Keeping the computers supplied
In a June 2023 conversation about AI’s electricity appetite, Wytock placed growing computer demand alongside broader electrification. His expectation was that several industries would require more electricity, increasing the importance of managing when it is used. He also allowed himself a small joke: “we work for the computers in many ways.” The line suits a former Google engineer whose company helps supply the machines.
That joke has become a business setting rather than merely a punch line. Gridmatic’s current focus includes renewable generation, batteries, flexible demand and customers such as data centers. Wytock’s stated purpose is to help renewable energy grow through better electricity-market decisions. The ambition requires more than an attractive prediction. It requires contracts, operating performance and customers whose demand continues through inconvenient hours.
The 2021 decision remains a useful episode to keep in view alongside those ambitions. A forecast represents a judgment about a future that has not happened. Operating a power business means accepting responsibility for what follows from that judgment. Wytock’s work lives in that interval, between a calculation and its consequences. Sometimes the next move is to charge a battery. Sometimes it is to wait.
“we work for the computers in many ways”Matt Wytock, June 2023