What Texas’s AI lab for 500 athletes means for the Army’s 450,000 soldiers

On Friday, October 23, Texas athletic director Chris Del Conte stood up at the Business of College Sports Summit and announced Bevo Lab, an AI and analytics platform that will serve all 21 Longhorn varsity programs and roughly 500 student-athletes, built with Dell Technologies, Google Cloud and SHI International and run out of the McCombs School of Business by Kirk Goldsberry, the former Spurs analytics executive (link). About 40 student researchers will run the modeling and data projects (link).

By Friday night, Texas fans had decided the AI called the fourth-and-2 that failed against Florida the week before. Del Conte had to clarify that the play was an example of what the platform could suggest, not what it did (link). So Bevo Lab was less than a day old and already getting blamed for a loss. Welcome to being first.

Full disclosure before I go further: Texas is a Teamworks customer, and has been one of the better-run athletic departments I've watched when it comes to actually using performance data rather than collecting it for the brochure. I'm not writing this because Texas is doing AI. Everyone in college sports says they're doing AI. I'm writing because of the order Texas did it in, because of what it cost, and because the Army is standing at the same starting line with a much bigger roster and a much smaller wallet.

What Texas actually built

Strip away the press release and Bevo Lab is three things.

The first is a warehouse. Texas calls it the Bevo Lakehouse, and its job is to consolidate game film and wearable data that previously lived team by team into one unified store, starting with football (link). That is not glamorous. It is also where most of the work is. Texas has spent years instrumenting its programs; the lab didn't create the data, it finally put it in one place where a model can reach all of it.

The second is a set of walls. Each team gets its own AI framework, which means Steve Sarkisian and women's basketball coach Vic Schaefer can't see each other's data or query the same platform (link). Goldsberry was blunt about why. There's a blurry line between sports analytics and medical data, these are young people, and he wants partners who already handle privacy for hospitals and the U.S. government so that "only people who have credentials can even see that data" (link).

The third is the part everyone writes headlines about: chatbots. But listen to how Goldsberry describes them. Each one is attached to "a pretty small amount of the world of data." His example is baseball. Texas is playing Vanderbilt for three games, the lab has every pitch Vanderbilt's starters have thrown, and instead of going through a spreadsheet with a highlighter the way he's done for 30 years, Jim Schlossnagle asks: what are this guy's top three pitches, how often does he throw the fastball, where does he miss?

That's the whole product. Not a dashboard. A coach asking a question in English about a curated slice of data and getting an answer he can act on before Friday.

Goldsberry has a line for the philosophy underneath it: "a strategy without data is a wish". I'd add the corollary he built the lab around: AI without a warehouse is a demo.

What it costs

Now the part the press release leaves out.

Texas Athletics spent $375.9 million in fiscal 2025, a record, on $352.5 million of revenue. That's $148 million more than Texas A&M spent in the same year, and more than Ohio State, Michigan or Alabama have ever reported (link). Spread across roughly 500 athletes, Texas runs its department at about $750,000 per athlete per year.

Bevo Lab sits on top of that. It has a business school attached, 40 student researchers who want the resume line, and three technology partners (Google Cloud's president is a Longhorn alum) who view a flagship sports AI deployment as marketing. Goldsberry said it himself: the lab wouldn't exist without the athletic director's personal insistence, and he can't walk into a football office without that.

None of that is a criticism. It's a description of a one-off. There are about 360 Division I athletic departments and only one of them is Texas. The next closest in terms of spending is Ohio State, and Texas outspent them by $50M.  Most of the schools in the Sun Belt, the MAC and the FCS run their entire department on less than Texas spends on football alone, and nobody at Dell is flying to Conway, Arkansas, to film a case study.

The Army is further down that curve than any of them. Its Holistic Health and Fitness program, the one that embeds dietitians, athletic trainers, strength coaches and cognitive specialists in combat brigades, costs about $3 million to stand up per brigade and $2.5 million a year to run (link). A brigade is around 4,000 soldiers. That's roughly $600 per soldier per year for the whole human performance apparatus, people included. Texas spends more than a thousand times that per athlete.  I’ll acknowledge that it may be an unfair comparison to hold up a budget that includes coaches salaries, athlete rev share, facilities, and everything else that goes into running the athletic program to just the thin slice of the Army’s budget represented by human performance, but the point still stands even if you look at the overall Army budget - with an FY26 budget of $197B, the Department of the Army had a budget allocation of ~ $435k per active duty soldier (and closer to $207k/soldier if you include National Guard and Reserves in the denominator). Texas Athletic’s per-athlete budget (before Bevo Lab) is still 70% - 350%+ greater than the United States Army’s per-soldier budget (link).

I want to be careful here, because this is where a defense vendor usually starts complaining about budgets. I'm not. The Army's number is the right order of magnitude for an institution that has to do this for everyone, forever, out of appropriated dollars. The point is simpler: whatever H2F's software looks like, it cannot be built the way Bevo Lab was built, because the Bevo Lab way requires Texas and Bevo Lab money.

The Army is building the same house anyway

In 2018 the Army piloted H2F in a handful of brigades. It worked well enough that in 2024 the Army said it would extend the program to every soldier, with 129 active brigades plus Guard and Reserve units equipped with performance teams by fiscal 2029 (link).

The software side of that is the H2F Management System (H2FMS), which the Army describes as delivering soldiers and leaders the kind of data-driven insights "typically only found within Special Operations or professional athletics" (link). The requirements call for one record spanning all five H2F domains, automated triggers when something goes wrong, and aggregated, anonymized views for leaders (link). If you squint, that's a lakehouse, a set of walls, and an interface. Same three things Texas built.

And the Army's leaders are worried about exactly the right problem. Speaking at a MOAA event in September, Command Sgt. Maj. Alex Kupratty of Army Transformation and Training Command warned about "data fatigue" as the dashboards come online, then offered the most honest sentence anyone in military human performance said this year: "I wear a Whoop. I wear an Oura Ring, and I tell you I probably look at half of the data that it provides me daily, and that's being very generous" (link).

Jonathan Flinn, who runs H2F for the 199th Infantry Brigade, who I’ve quoted before, put the stakes plainly after evaluating three candidate apps at Fort Benning: previous data efforts at the 75th Ranger Regiment and elsewhere "failed because we got to get the end user buy-in." A second pilot focused on the app and end-user experience is planned for the second quarter of fiscal 2027.

Those are the correct worries. Texas answered them in public. It just used 40 students and a hyperscaler to do it.

Three things Austin figured out

Coaches ask questions. Texas didn't ship the Vanderbilt dataset as forty charts. It shipped a box you type into. The Army equivalent isn't a sergeant major scrolling heart rate variability for 800 soldiers. It's a first sergeant asking, "Who in my company shouldn't ruck Thursday, and why?" and getting a short answer with reasons attached. The H2FMS requirements already describe the ingredients, triggers and leader roll-ups. The question box is the interface that makes those ingredients usable by someone who did not go to school for statistics, which, as Goldsberry notes about coaches, is nearly everyone who matters.

The walls are part of the product. Texas decided on day one that football can't see basketball, and picked infrastructure partners already trusted with medical and federal data. For a university this is prudence. For the Army it's the price of admission. A soldier's sleep, nutrition and injury history is about as sensitive as personal data gets, and it lives in a system that has to clear accreditation gates most sports vendors have never heard of. The lesson from Austin is that security isn't the step after the fun part. It's the thing that let the fun part exist.

One record per human comes before any intelligence about that human. The Bevo Lakehouse exists because film and wearables used to live in separate silos, one per sport. H2F's five domains, physical, nutritional, mental, spiritual and sleep, have historically lived in separate places too, some of them paper. A recent Front Office Sports survey, co-authored by Teamworks, of 200 sports executives found that organizations with athlete data in one to three systems reported clear returns from AI 22% of the time; at four or more systems, that dropped to 8%. Connecting existing systems was the single most common planned AI investment, at 58% (link). The Army doesn't need a survey to know this. 

Bevo Lab for everyone else

Here's where I stop describing and start declaring an interest, because this is the problem Teamworks exists to solve, and the next version of it is about to ship.

Texas built its warehouse by hand because its data lived in a dozen places (including some Teamworks software). For most of the 7,000-plus teams that run on Teamworks, the warehouse already exists: it's the Teamworks system of record their staff log into every day, with every athlete's identity resolved across scheduling, nutrition, strength, medical and communication. Our AI runs inside that system, on that record, and nowhere else. Your data is never sold, never exposed to another organization, and never used to train a general-purpose model (link). That's the lakehouse and the walls, and you didn't need a business school and a hyperscaler partner to get them.

What's new is the third piece, and it's arriving in stages because that's how trust gets earned.

The first stage is live now with a subset of our customers: a question box inside the product, scoped to your own data, that answers the kind of thing Schlossnagle asked about Vanderbilt. A dietitian asks which athletes on the travel roster haven't logged a meal in three days. A clinician dictates an injury note and the required fields fill themselves for review. The answers cite their source and the model says so when it can't help.

The second stage is agents that notice repeated work and offer to take it over. The weekly readiness report someone rebuilds by hand every Monday. The threshold breach a coach catches on Wednesday that should have been flagged Sunday night. The agent builds it, keeps it current, and a person approves before anything leaves the building.

The third stage is where Bevo Lab is pointed and where we're headed: one request that draws on several domains at once. Adapt this soldier's training block for his injury status and what he's eating, and give me one recommendation I can approve. Change it in one place and the others follow.

I’m confident that I’m not overstating my case when I say that no other technology provider has the ability to provide this capability as cleanly or as capably as Teamworks.  The breadth of our specialist software supporting teams across operations, recruiting, finance, and human performance, as well as our expertise integrating with third-party systems, places us in a unique position to deliver on the promise of AI.

Notice what isn't on that list. No student researchers. No hyperscaler partnership announced at a summit. The design is the same as Texas's. The delivery is what changes when you have to serve a P4 customer who’d rather put the extra tens of millions of dollars spent on a bespoke lab into other priorities, a Sun Belt school with a four-person performance staff, or an infantry brigade at $600 a head, instead of one department with $376 million.

And the military version isn't a low-effort attempt to resell existing software into a new market. It runs in the government cloud environment the Army requires, under the security controls the Army specifies, because the walls Goldsberry talked about are a nice-to-have in Austin and a condition of entry at Fort Benning.

The shape of the thing

The Department of War has already said where it's going. The Warfighter Performance Optimization memo signed in May commits to "aggressively fielding data analytics, wearables, and cutting-edge technologies," and admits the department maintains its weapon systems with more rigor than it manages the people operating them (link). A companion directive gave the services a June deadline to report, among other things, how they're using wearables and mobile sensors (link).

So picture H2FMS a few years from now, shaped the way Texas shaped Bevo Lab but paid for the way the Army pays for things. A soldier opens one app and sees three things, not thirty, because somebody decided what matters and hid the rest. A first sergeant types a question about Thursday's ruck and gets an answer with names, reasons and a recommendation. A brigade surgeon sees the trend line across 4,000 soldiers without seeing a single name. A division commander asks whether the H2F investment is paying off and the answer is a number, with the arithmetic shown. And under all of it, one record per soldier, walled off by role, sitting in infrastructure that was accredited before the first question was ever asked.

That’s the house Texas built for 500 people with 40 students and some very generous partners. It’s the house the Army needs for 450,000 people at a brigade’s budget. And the first time a ruck march goes badly, somebody is going to say the AI told them to do it, the same way Texas fans decided the AI called fourth-and-2.

That’s fine. That’s what being first costs. The next morning, the first sergeant can ask what actually happened, and the right H2FMS will tell him.