← Field Notes

From 250 Pitches to Fusion Power: What AI Still Can't Replace

On Episode 4 of Volume Control, Voxus managing partner Kevin Pedraja tells Natalie Luke and me why AI can't write the pitch a reporter actually opens, why electricity is becoming the real limit on AI and why he's still excited about fusion, caveats and all.

Viet Nguyen headshotViet Nguyen

Kevin Pedraja has a name for the gear you need to work in tech PR. "You have to strap on your rejection immunity suit and just keep coming to work every day and keep doing it," he told Natalie and me on Episode 4 of Volume Control.

Kevin is a managing partner at Voxus, a tech PR agency, and has been in the business since the tech PR boom of the mid 1990s. He gave us the math behind the suit. "I'll reach out to a reporter who's gotten 250 pitches that day. So I have to come up with a really good one to get his or her attention."

That's the part of his job people keep trying to hand to AI, and it's the part AI is worst at. The rest of the conversation went somewhere I didn't expect. We started with chatbots and ended up on substations, power bills and the long bet on fusion. The thread connecting all of it: as AI models get cheaper and more alike, the scarce things are judgment and electricity. Both come back to people. Who do you trust to tell the story, and who pays for the power?

Where AI actually helps a PR team

Kevin isn't anti-AI. His team uses it to learn a client's technology fast, and he treats it like an assistant for tightening his own writing. What he won't do is let it write the first draft. "They're the sum total of a lot of knowledge, but that means they're very kind of middle of the road," he said. "So when you're creating content, you don't want to sound like everybody else."

His best example is research. A model can read a reporter's last 50 stories and tell you what they have in common, which helps you figure out how to approach a topic with them. It can't write the pitch. "It's not like, give me a pitch for this reporter based on what they've written. A lot of people are trying to sell that product. I don't know that that product really works." Reporters can spot an AI pitch, he said, and they complain about them constantly.

I've spent most of my career in tech communications, and this matches what I've seen. When your name or your company's name goes on something, it's your credibility at stake. A reporter who gets 250 pitches is filtering for one thing: does this person know what I cover and have a real story? A model that averages the internet is built to produce the opposite.

Replacing people is backfiring

Kevin's read on the last two years is blunt. "I think some companies over indexed on replacing people rather than helping people be more efficient and be better at their jobs."

Meta is the loudest example so far. Reuters reported in August that the company put tracking software on U.S. employees' devices to capture keystrokes and mouse clicks, so AI agents could learn how people use computers (Reuters). Favorable sentiment on Meta's internal employee survey fell from 74% to 55%, and executives paused the mouse-tracking program.

There's also a money problem nobody budgeted for. Kevin pointed out that costs are already climbing, "and that's changing, I think, the perception of what AI can do for companies, because no one wants to jump in and think they're gonna save a lot of money by replacing people, only to turn out that their token cost is higher than their actual employee costs."

That's the AI meter showing up again, which I covered in my last Field Note, AI is getting a meter. Someone has to read it.

Today's breakthrough, tomorrow's commodity

Kevin has watched a few hype cycles. "One of the things I've learned is that technology is more incremental than it is revolutionary," he said. "There's very few things that are revolutionary. And even when they are, they're not revolutionary overnight, they're revolutionary over time."

His example was models being open sourced in China that are "performing almost as well as these really expensive models." "What's revolutionary one day becomes a commodity the next day," he said. "That is the fundamental law of technology."

If that's right, and I think it mostly is, the advantage moves away from the model and toward the things you can't download. Judgment is one. Power is the other.

Every model runs on the grid

Every chatbot answer, every agent and every image runs on a chip in a data center that has to be plugged into something. The International Energy Agency (IEA) estimates data centers used about 485 terawatt-hours of electricity worldwide in 2025 and projects about 950 terawatt-hours by 2030 (IEA). In its earlier report, the IEA said that 2030 figure is slightly more than Japan's total electricity use today (IEA).

Kevin was careful not to blame it all on AI. "AI data centers are definitely kind of supercharging that demand, but they're not the only culprit," he said. Population growth, electric cars and all the screens in our houses are pulling on the same wires. "We've electrified a lot of things," he said. We recorded on a day when it hit 91 degrees in Seattle, which he called "unheard of not too long ago."

Then there's the grid itself. "The grid itself is old and it's not, its capacity has not kept pace with demand," Kevin said, "and so you know there's gonna come a reckoning at some point where we're gonna have to spend a lot of money on just the grid."

Seattle is already making that call. In June, the City Council voted 9-0 for an emergency one-year moratorium on new data centers that need more than 20 megavolt-amperes of power (Seattle City Council). The council acted after four companies proposed five data centers that together could draw 369 megawatts, which the council says is roughly enough for 300,000 homes. Seattle City Light is now writing rules for how large new loads connect and who pays for it (Seattle City Light).

Seattle's moratorium isn't about Microsoft, by the way. Microsoft's Washington data centers are on the other side of the mountains, in Quincy, East Wenatchee and Malaga (Microsoft). That points to the bigger Washington story. About 74% of Washington's electricity was generated east of the Cascades in 2025 (Energy Information Administration), while 78% of us live west of them (Office of Financial Management). The lines that carry power over the mountains are full, and the next 500-kilovolt line upgrade across the Cascades isn't due until 2032 (Bonneville Power Administration). Here's what that looks like on a map.

The power is east. The people are west. A map of Washington split at the Cascades. West of the mountains: 78% of residents and 26% of power generated. East: 22% of residents and 74% of power generated, according to the U.S. Energy Information Administration for 2025. Power lines cross the Cascades from East Wenatchee, Malaga and Quincy toward Seattle. West: Seattle saw five proposed data centers seeking 369 megawatts, and the City Council paused new large ones for a year, 9-0, in June 2026. Across the mountains: 2032 is when the Bonneville Power Administration's next 500-kilovolt line upgrade across the Cascades is due. Puget Sound Energy estimates it needs 3,000 to 4,500 megawatts more capacity across the Cascades, which means at least two new 500-kilovolt lines expected in 2035 or later. East: data centers used 37% of Grant County's electricity in 2025, with eight operators and about 800 megawatts more in large-load requests waiting. Microsoft's Washington data centers are in Quincy, East Wenatchee and Malaga. Who pays for the wires: Microsoft paid $86.5 million up front for the Jumpoff Ridge substation serving its Malaga data center, and average Bonneville transmission rates rise 19.9% for 2026 to 2028.

My view: Washington needs more transmission across the Cascades. Wires alone won't fix it, though. In 2025, data centers used 37% of Grant County's electricity (Grant PUD), so the east side's surplus is shrinking before it ever reaches the mountains.

Who pays for it

This is where Natalie said what a lot of homeowners are thinking: it would be nice if some of those billions came from someone like Microsoft instead of her.

Kevin answered as a ratepayer. "You and I as individual ratepayers, we don't want to see our electricity bills go up 15 to 20 to 30% so that a large company can make more money," he said.

Some of the big buyers are now saying they'll cover it. In January, Microsoft President Brad Smith wrote, "We'll pay our way to ensure our datacenters don't increase your electricity prices," and said Microsoft will pay for transmission and substation upgrades its data centers require (Microsoft). Microsoft has also signed a 20-year deal to buy power from Constellation's restart of Three Mile Island Unit 1, now called the Crane Clean Energy Center, which Constellation expects back online in 2028 (Constellation).

Closer to home, Chelan County's public utility says Microsoft paid up front for the $86.5 million Jumpoff Ridge substation that serves its Malaga data center (Chelan PUD). Everyone else pays through rates. The Bonneville Power Administration, which runs most of the Northwest's high-voltage lines, raised its average transmission rates 19.9% for 2026 to 2028 (BPA).

Utilities have a second problem that has nothing to do with AI. Kevin described it this way: the more people with money install rooftop solar and use less from the grid, "the more the burden shifts to people who don't have the means to pay." The fixed cost of the wires doesn't shrink when the wealthier customers leave.

Building new supply is slow, too. Kevin said connecting a new power plant to the grid can take "five to ten years or longer." Lawrence Berkeley National Laboratory found that for projects built in 2025, the median time from interconnection request to commercial operation was over five years (Berkeley Lab). That's a big reason the large AI companies are building power "behind the meter," with generation wired straight to the data center instead of waiting on the grid.

That approach has its own fights. In Memphis, xAI ran gas turbines to power its Colossus data center, and the Southern Environmental Law Center says as many as 35 ran without permits before the county permitted 15 (SELC). The backlash has grown as the company expands (Scientific American).

Kevin didn't pretend to have a fix. "There's not a simple answer to any of them," he said. "They're all really complicated. And there's a lot of interdependencies that are hard to solve."

Fusion, with the caveats attached

The hopeful part of the episode was fusion, and I'll be upfront about why Kevin knows it well. Disclosure: Voxus represents the local fusion energy company discussed in this episode. Kevin called himself "a total geek and science fiction nerd" and said working on a fusion company was a dream come true. He was also careful to say he didn't want to speak too much for the company.

So here's where fusion actually stands. The Nuclear Regulatory Commission says there are "no fusion machines that are commercially operational in the United States that are providing electricity to the grid," though more than 50 companies around the world are working on it (NRC).

Kevin brought his own skepticism, which I appreciated. "I always heard fusion was 20 years away, and it's been 20 years away for the last 50 years." Every fusion date you read is a target, not a promise.

I'll admit my nerd brain went straight to the endgame on the show. Put a fusion generator next to a data center and it's just a funnel. Hydrogen goes in one end and tokens come out the other. That's a long way off, if it happens at all. Here's what that funnel would look like, and how much of it is still on paper.

Fuel in. Tokens out. How a fusion plant could power AI. It's still years away. Five steps on one pipeline. Fuel: forms of hydrogen and helium; there are 33 grams of deuterium per cubic meter of seawater. Fusion: atoms fuse and release energy at over 100 million degrees Celsius. Not yet: no U.S. fusion plant sends power to the grid. Grid: utility power lines; 74% of Washington power was made east of the Cascades in 2025. Data center: runs the AI; one AI-focused data center draws 100 or more megawatts. Tokens: the AI's answers; a typical Gemini text prompt uses 0.24 watt-hours, per Google. One 50-megawatt fusion plant equals about one-sixth of the 288 megawatts Microsoft's Malaga data center is authorized to use. Disclosure: Voxus represents the local fusion energy company discussed in this episode.

The power bill is due now.

Back to the 250 pitches

The part of this conversation that stuck with me most wasn't the physics. It was Kevin describing what good reporters do. "The best reporters I know actually get out into the world and they talk to people and they go to events and they develop sources," he said. "You can't replicate that with an AI."

The same is true on his side of the inbox. A model can tell you what a reporter wrote last month. It can't tell you whether your story is worth her time, and it can't make her trust you. Kevin's whole business is the relationship that gets the email opened.

You can hear the full conversation with Kevin on Episode 4 of Volume Control.

Get the next one

Field Notes, once a month, in your inbox.

What we are seeing across AI work with organizations in the Pacific Northwest, plus a short email when a new Volume Control! episode drops. No vendor pitches, no daily drip.

We confirm by email first. Opt out with one click, any time. Privacy

Written by

Viet Nguyen headshot

Viet Nguyen

Co-founder, Third Space Project

Viet runs advisory at Third Space Project and co-hosts Volume Control. Based in Seattle. Writes about what he is seeing across the work.

Email updates

Get Third Space updates.

A monthly note on what is working in AI adoption, plus a short email when a new Volume Control! episode drops.

We confirm by email first. Opt out with one click, any time. Privacy