Back to the episodes

Kanav: fourteen posts

Fourteen posts in your voice, each built from something you said on The Punch List or on stage. Tap a first line to read the post. Under each one: where it came from, and the one thing we'd check with you before it goes out.

1Jensen Huang said last year it takes about three years to build a data center in America. I know three general contractors who built theirs in 11, 12 and 14 months.

Jensen Huang said last year it takes about three years to build a data center in America. I know three general contractors who built theirs in 11, 12 and 14 months.

You'd guess they had more workers than everyone else. If anything they had it harder, because a data center needs a lot more electricians than most buildings, and there aren't enough to go around.

What they asked for was control: "Give me control over design. And give me control over subcontractors." Instead of bidding out every trade five ways, they brought in subs they had worked with before and already knew how to coordinate with.

So most of the speed was decided before anyone broke ground. One team owned the drawings and the people building from them, so when something didn't fit, the people who had to live with the answer could settle it quickly.

On most buildings, that same problem has to travel between an architect, a few engineers, a general contractor and a long list of subcontractors, and every answer reopens someone else's plan. I'm an owner on two projects myself, and every change I ask about takes six to eight weeks to come back.

At MeltPlan, we're building towards letting any team check a change against the code, the cost and the schedule in a day instead of six weeks.

Whatever we're learning from building data centers this fast should be used on every other kind of building. We need to ship things fast.

From your Punch List episode on China and speed (22 Oct): the three GCs at 11, 12 and 14 months.

To check: Happy to tell the three-GC story on LinkedIn, and are both your owner projects still live?

2Almost every owner I meet has a story about a contractor who went over budget. Very few of those stories start with the contractor.

Almost every owner I meet has a story about a contractor who went over budget. Very few of those stories start with the contractor.

I had one too. When my wife and I remodeled our house in Berkeley, the budget and the timeline both ran 50% over. I thought I'd been unlucky, so I asked around. My friends told me I'd done fine.

That's when I started digging. Most budgets get set before anyone has the full picture, on drawings that aren't finished and will change. Then every change has to travel through the architect, the engineers and every trade involved, and each one adds cost the first number never had.

So by the time the contractor is building, a lot of the overrun is already baked in. My co-founder TK ran construction projects for years, and he puts it bluntly: "You can screw things up in ops, but you're never gonna perform better than what pre-construction gave you."

Contractors do owe owners one thing, and that's the honest number before anyone signs, even when a lower one would win the job. When I was in Dubai this year, everyone in the industry seemed to know the early budget would end up around 50% over, and nobody wanted to be the first to say so.

That's why we built MeltPlan around the planning, before anyone breaks ground.

So if I'm the owner, I'd rather hear the honest number before I sign and spend the extra weeks on the plan than blame the builder halfway through.

From your Berkeley remodel, TK on pre-construction, and your read of Dubai.

To check: The remodel ran 50% over (you told EUVC 40%). Which is right, and fine to mention your wife?

3"When are you leaving tonight?"

"When are you leaving tonight?"

That's what our hospital customer in Iowa asked when my team flew in, back when I was building my last company, Innovaccer. They had just seen our first prototype, hated it, and given us three months to fix it.

I told them we were staying. We camped there for those three months. One night the police came into the administrative offices at 4 a.m. because they couldn't work out why the lights were on, and I had to explain that I was working there.

When they first asked us to build it, we had no idea what it even meant. In those three months we learned it in their office, five layers deep, and by the end they offered to set up an automatic monthly wire to us instead of waiting for our invoices.

They were surprised anyone stayed. The biggest health systems get vendors flying in all the time. Who shows up in Des Moines and stays?

I think the same is true in construction today. The middle of the market, the thousands of contractors outside the top 400, rarely gets anyone who sits with their real problem. So that's where we spend our time at MeltPlan, and before a first call I ask contractors to send us drawings from one of their past projects.

I still think the most important thing in selling to businesses is being honest with yourself about whether you really understand the customer's problem.

From your Innovaccer years: Lightspeed and TiECON.

To check: Fine to name Innovaccer and keep the Iowa hospital unnamed?

4The things that go into new hospitals, clinics and other health care buildings cost 12.1% more than they did a year ago. What builders charge for them went up 6%.

The things that go into new hospitals, clinics and other health care buildings cost 12.1% more than they did a year ago. What builders charge for them went up 6%.

Those are the government's price numbers, out last week. Yet the latest contractor survey found builders more hopeful about their profit margins than a month earlier.

If you signed a fixed-price job last fall, you're now buying this year's steel and copper out of last year's number. Steel mill products are up 32.9% and diesel 115%, and that gap comes straight out of the margin.

Back in May, I said every pre-con room would start adding 10% for unknowns. The trouble is that prices haven't moved evenly. Ready-mix concrete is up 2.4% and gypsum is actually down.

A blanket 10% overprices a lot of your bid and may still come up short on the steel. And when more than half of contractors had a project cancelled, postponed or scaled back in the six months to August, a padded number is an easy way to lose the job.

So I'd be surgical about it. Look at what's heavy to ship, what's imported and what takes a lot of energy to make, and put the padding on those lines.

I'd want every escalation allowance in a bid to name the material it's for.

From the November producer price numbers and what you said on The Punch List in May.

To check: "Out last week" needs the date of the next release before it goes out.

5Owners used to read a contractor's bid and say, "Sounds good, go forward." Now a lot of them ask Claude first.

Owners used to read a contractor's bid and say, "Sounds good, go forward." Now a lot of them ask Claude first.

I hear this a lot from general contractors and architects. The GC buys out the subcontracts, levels the bids and sends the owner a clean report. The owner pastes it in and asks, "Hey Claude, what do you think?" Ten new questions come back, and the GC spends three weeks answering them.

I've done it myself. We built a small 8x10 gym in our backyard, and my contractor's quote was one page with about ten bullet points. I put it into Claude, and it told me to make sure the paint was included. I'm not a builder, so I hadn't noticed it was missing. When I asked him, he said, "Oh yeah, good point. Let me add that." Otherwise it would have come back to me later as a change order.

The old way made sense. Most owners had no practical way to check a bid line by line, so they trusted the people who did this every day.

That check now costs an owner almost nothing. My co-founder TK says there are fewer places to hide, and that lets competence shine through.

So the bar for a bid has moved. Write the scope out in full, say the exclusions plainly, and be ready to explain every number, because someone is going to ask.

That's hardest on the contractors who leave gaps in the scope to fill later as change orders.

From your Punch List episode on Anthropic's S-1 (6 Nov): owners asking Claude, and your backyard gym.

To check: Are you happy to tell the gym story, and is "owners now ask Claude first" what you keep hearing?

6Everyone is telling construction teams to use AI to read drawings and count quantities faster. That helps, but it doesn't touch the reason pre-construction takes so long.

Everyone is telling construction teams to use AI to read drawings and count quantities faster. That helps, but it doesn't touch the reason pre-construction takes so long.

Software developers have something called a compiler. When two engineers put their code together, the compiler tells them in seconds whether it runs. If it doesn't, they fix it and try again.

Pre-construction has nothing like that. When an owner changes the design, someone has to check it against the city's code, someone has to price it and someone has to schedule it, and each of them needs days.

My co-founder TK lived this as a general contractor. When he ran pre-construction for a developer, Jon Wright, Jon would come in with a new idea every few days and ask for a price and a schedule. TK's estimator had to look at it, then his superintendent, and every change to the design needed more time.

So teams try fewer versions. The design that gets built is the one there was time to check, and the questions nobody had time to ask turn up during construction, when someone walks in and says, "Why didn't we think of this?"

Faster takeoffs are still worth having, and we sell them too. On their own, they speed up one step of a check that still takes weeks.

What pre-construction needs is four compilers: is it allowed under the code, what will it cost, how long will it take, and can it be done better. That's what we're building at MeltPlan.

If you're deciding where AI goes first, start by counting how many versions of a design your team can check in a week.

From the same episode: "precon needs compilers", with TK's story of pricing Jon's changes.

To check: Fine to name Jon as TK's client, and to say the full code, cost and time check is where MeltPlan is headed?

7The usual explanation for why data centers go up so fast is money. Money helps, but I think the bigger reason is that nobody changes the floor plan.

The usual explanation for why data centers go up so fast is money. Money helps, but I think the bigger reason is that nobody changes the floor plan.

A data center owner's requirement fits in one sentence: I need 50 megawatts of the latest chips, and I don't care how you do it. Once that is set, the owner stays out of the design.

Here's the test I'd use on any project. After the owner tells you what they need, come back three months later with a floor plan. Will they change it?

In a house, they will. In a hospital, they will. In a data center, they barely look at it, because the only question is how many chips fit.

Each change matters because one change pulls others with it. A wall that moves takes ductwork, wiring and sometimes a fire exit with it, so the drawings, the price, the schedule and sometimes the permit get done again.

To be fair to hospitals, those changes usually come for good reasons. As my co-founder TK says, wherever a lot of people will use a building, they change their minds as it becomes real, and someone walks in and asks for a window. A data center only changes when the chips change, and the owner knows about that well in advance.

So on a hospital, I'd get the changes out early and run as many versions as possible in pre-construction, while a change is still a drawing.

Before you promise anyone a fast schedule, run the floor-plan test on your own project.

From your floor-plan test on The Punch List (6 Nov), with TK on why people change their minds.

To check: "The usual explanation is money" is our framing. Would you say it?

8My bet for next year: America won't add the 40 gigawatts of data centers it planned. It could be closer to 20, and not because the demand went away.

My bet for next year: America won't add the 40 gigawatts of data centers it planned. It could be closer to 20, and not because the demand went away.

The US has about 40 gigawatts of data center capacity today, and the plan was to add roughly 40 more next year. Doubling that fast needs power, grid connections and permits, and that needs a state pushing from the top.

The signals are going the other way. On Tuesday, Loudoun County in Virginia, the biggest data center market in the country, voted to pause new data center zoning requests for up to a year. Texas's governor-elect has promised a moratorium on new data center permits. Even Anthropic's IPO filing warns that opposition to data center development "has also become more organized and coordinated across jurisdictions."

What's missing is a push from the top, across the country, saying let's build this. Without it, the same buildout could take four years instead of one.

I think the owners who move first will pay for a faster yes. Say a data center costs $15 million a megawatt. If a governor offered $20 million a megawatt, with the extra $5 million going into parks for the town and a promise you can build in one year, AI owners would take it happily.

If you're a contractor with data center work in your backlog, plan for slower approvals next year and take what you learned about building fast to hospitals, plants and schools. If you're an owner, put the price of a community's yes in your budget now.

From your midterms episode (19 Nov), the Loudoun vote (17 Nov) and Anthropic's own filing.

To check: The 40 GW numbers and the $15M-to-$20M a megawatt idea: yours, and still how you see it? "On Tuesday" becomes "On 17 November" if it goes out later.

9What I call the shadow AI committee is showing up in the construction companies we sell to.

What I call the shadow AI committee is showing up in the construction companies we sell to.

It's a group of people who decided on their own to test AI tools for the company. Nobody appointed them, and there's no title on LinkedIn that says "AI committee", so you won't find them on an org chart.

On sales calls, buyers tell us we're the tenth startup through their door, and that they've tried five demos in the past five months and all of them failed.

Some of these companies have no AI strategy at all, but they still have a committee. People raise their hands and take on evaluating startups as a second job.

Compare that with how a contractor buys almost anything else. A new crane or a new accounting system has an owner, a budget and someone who signs off. AI, which could change how every estimate and bid gets made, is often being decided by volunteers in their spare time.

These people are doing their company a favor. They're curious, and they care. But a second job gets second-job hours.

The few mature companies we meet work differently. They hand us their own test, with their own data, and expect the answer back in one to four days.

So find out who is on your shadow committee. I tell the buyers we meet to have a test ready from one of their past projects, with the data masked if IT insists, before the next vendor walks in.

From your a16z Tech Week panel (9 Oct): the shadow committees and the five demos in five months.

To check: Is "shadow AI committee" a name you'd put your weight behind?

10In March 2020, at Innovaccer, the healthcare company I co-founded, we got a message from regulators that came down to one word: build.

In March 2020, at Innovaccer, the healthcare company I co-founded, we got a message from regulators that came down to one word: build.

We made software for hospitals and health systems. When the lockdown hit, we realized almost overnight that our customers needed telemedicine. Virtual care was less than 4% of care in the US at the time.

Then the regulators said they were dropping the rules on telemedicine. We built our telemedicine software in about seven days, and the whole industry built with us.

But we were still liable if we leaked a patient's health data, and the safety rules stayed. What went was the paperwork around them, like the permissions and recordings that slowed everyone down.

I think construction needs the same sorting. Almost every rule gets put under the umbrella of fire and life safety. Some of those rules keep people alive, and some are just control.

The host of our show The Punch List, Jon Wright, builds modular single-family homes. After the city issues the master building permit, each house still goes through 24 permits or inspections during construction. If he built the same house on site, he says it would be closer to 32 or 34. I can't tell you which of those are needed.

If you build or approve buildings, take the permit and inspection list from your last project this week and mark each line: does it protect someone's life or health? Keep every one that does, then ask who still needs the rest.

From your Punch List episode on China and speed (22 Oct), with Jon's 24 inspections.

To check: Was the first telemedicine app really about seven days, and is Jon fine being quoted?

11An AI agent is not a product. Most of the build-or-buy argument about AI right now treats them as the same thing.

An AI agent is not a product. Most of the build-or-buy argument about AI right now treats them as the same thing.

An agent does one task when you ask it, like leveling these bids or drafting these RFIs. A product holds the work around that task, every step people take before and after it.

Bid leveling shows the difference. Lining up the subcontractors' numbers is one step. Around it, you plug temporary numbers for missing scope, go back and forth with the subs, look across the whole list of subcontractors and award the work. You need something to prepare the bids, something to level them and something for everything after.

A second-generation contractor in Texas found this out a few months ago. His father started the business, and he's taking it over. He vibe-coded his own bid-leveling tool in Claude Code and got some results. When he wasn't happy with them, he asked Claude how to improve it, and Claude asked whether he'd tried vendors like MeltPlan. That's how he found us, and he uses our product now.

He was happy with the agent some of the time. What he couldn't get was the rest of the workflow.

Agents still matter. A good product will be full of them, and for a one-off question an agent on its own is often enough.

So when you're deciding whether to build an agent or buy software, ask what happens after the agent answers: who changes the number, who approves it, and where it's kept for the next bid. If nobody can say, the agent has done one step and the rest of the work is still on your team.

From The Punch List (1 Oct) and the a16z panel: the Texas contractor Claude sent your way.

To check: "For a one-off question an agent on its own is often enough" is our line. Agree?

12Every architect should ask the same question about AI: how do you know it's right?

Every architect should ask the same question about AI: how do you know it's right?

So we gave it the exam building inspectors take.

Building inspectors in the US have to pass an exam to get certified, and a few firms sell practice tests for it. We collected about 3,000 of those questions and had our code research system answer all of them, again and again, as we built it.

When we started, it got about 70% right. Today it gets more than 95%.

Give the same questions to GPT, Gemini or Claude and they might get 80% or more right. But they rarely back the answer with a good source, the section of the code it comes from.

For an architect, that's the part that matters. An answer with no source still has to be looked up by hand, so it saves very little time, and a wrong answer with no source is how people stop trusting AI.

At 95%, the first answer is usually right, and the source is right there, so a person can check it in a minute.

If you're testing any AI tool on building codes, ask for two numbers: how many answers it gets right on a test like this, and how many come with a source you can open.

From your talks in March and April: the 3,000 inspector-exam questions.

To check: Did you start at 70% or 75%, and is 95% plus still today's number?

13In our first two months of selling MeltPlan, an architect bought a 12-month subscription without talking to a single person on our team.

In our first two months of selling MeltPlan, an architect bought a 12-month subscription without talking to a single person on our team.

The usual line is that construction doesn't buy that way. It buys from people it knows, after a handshake and a lunch. That advice made sense, because a wrong tool on a building costs real money and nobody wants to be the first to try something new.

Our AI agent wrote to her first. She had a question, and the agent answered it. She signed up herself and started using our code research product. We give a one-month trial, and she used it for 3 weeks. Then she put in her credit card for the full year.

Nobody from marketing or sales touched it. It was a very small sale, but it was our first one with no human in it.

I think it worked because the product answered her real questions in the first three weeks. She could see the answers for herself.

The usual advice still holds for big deals. Large contractors and architecture firms now want it across their whole portfolio, and that's why we hired experienced enterprise account executives this year.

But I'd make the same choice again: let the people who want to try something on their own do it, all the way to paying, and keep our account executives for the big deals.

From your April interview: the architect who bought a year with no human in the loop.

To check: Is "construction buys from people it knows" the advice you'd push back on?

14Most firms hand AI to their youngest people first. The best user I saw this year was the oldest principal at the table.

Most firms hand AI to their youngest people first. The best user I saw this year was the oldest principal at the table.

This spring I had lunch with a team at one of the largest architecture firms in the US. We talked about our product and showed it to them.

After lunch, the oldest principal in the firm came over. He said they'd buy a product like this someday, after a few more hoops. But he had a client query he'd been stuck on for three weeks, and he asked if we could try it right then.

We ran it, and the query was answered in about five minutes. It mattered, because a wrong answer to that question would have meant a big extra cost in the build-out. I could see the spark on his face.

People like him get left out of AI rollouts for understandable reasons. Many of them have tried ChatGPT or Gemini, got made-up answers and lost trust. And nobody expects the most senior person in the firm to run a pilot.

But they're exactly who AI needs now. It's right most of the time, and the hard part is knowing when it isn't. Someone who has answered these questions for a whole career already knows the answer key.

They also know which questions are worth asking. He brought the one problem that had cost him three weeks.

So if you're starting an AI pilot, put your most senior people in it, give them their hardest open question, and let them judge the answer.

From the same interview: lunch at one of the largest US architecture firms.

To check: "Most firms hand AI to their youngest people first" is our framing. Does it match what you see?