General Motors says it simulated more than 300,000 designs for its next EV battery before building a single physical cell. The run took 300 million CPU hours, or one computer running around the clock for 34,000 years, by GM’s math.

GM pitched the number Wednesday at its GM Forward event as proof it can halve the traditional four-to-five-year vehicle development process, to about two years. With the battery, the team went from first pack concept to formal program approval in one month.

Nearly every announcement ran on the same logic. GM’s self-driving software logs more than 3 billion simulated miles a day, against about 1 million real-world miles a month. Factory teams rehearse assembly in digital models.

Fortune reported Tuesday on the culture change gripping the top American auto-seller: engineers rewarded for missed stretch targets and a battery chief who says, “I want to see red.”

Sterling Anderson, the former Tesla Autopilot chief GM hired as chief product officer in 2025, recently told Fortune it’s “effectively the third phase in the evolution of product development for humanity,” built on “probabilistic, large data-requiring methods.” This is what he meant. That said, the battery won’t reach volume production until 2028.

A battery built on a long shot

The cells use lithium manganese-rich chemistry, or LMR, which GM says offers 33% more energy density than lithium iron phosphate, or LFP, the cheap chemistry China dominates, at comparable cost. They use roughly 30% to 40% nickel, battery engineer Andy Oury told Fortune, versus about 85% in today’s high-nickel cells. GM plans to build them with LG Energy Solution at their Ultium Cells joint venture in Spring Hill, Tenn.

“We’re not licensing this from China,” Kurt Kelty, GM’s vice president of battery and sustainability, told Fortune. By his own account, most of the industry considers LMR a long shot, and GM made the call with perhaps 60% to 70% of the data. LMR has historically struggled with short cycle life and voltage decay, but GM believes it has solved that.

Proof will come from GM’s $900 million Battery Cell complex in Warren, Mich., due to be fully operational later this year and ramping toward a couple thousand cells a day.

The cone course

On Fortune‘s recent visit to Warren, Jason Fischer, a 28-year GM veteran who leads the AI push in engineering, demonstrated a digital car sliding off a cone course until engineers tuned it in simulation, work he said would take months with a real car. GM holds weekly “tall pole” reviews of whatever is slowing each job. “I’m convinced that we’re going to be able to do things just as fast as China,” Fischer said.

At GM Forward, the company said its fourth-generation EV platform should deliver about 20% more range per kilowatt-hour. It is expected to be 25% more aerodynamic and 15% lighter, with a 20% more compact pack and drive units with 40% fewer parts. In factories, GM’s PlantBrAIn platform has moved battery packs through assembly 30% faster, and Gen 4 vehicles will check their own wiring on the line. The platform will also carry an updated version of Super Cruise.

The brain goes into the Silverado

GM’s new central computer, with 40 times today’s processing power and 30% less wiring, debuts in 2028 on the Cadillac Escalade IQ, then spreads to GM’s highest-volume gas SUVs and trucks, three-row crossovers and Gen 4 EVs. It enables the other headline: autonomous highway driving, announced last year for the Escalade IQ, is coming to gas-powered full-size SUVs and light-duty pickups. Pre-production vehicles are testing on closed courses.

The driving software trains largely through self-play, with versions of itself playing the merging car and the surrounding traffic. Much of the method came from Cruise, the robotaxi unit GM absorbed in 2024. Rashed Haq, GM’s vice president of autonomous vehicles, told Fortune that GM kept “a lot of the methodologies” for the next-generation Super Cruise. Current Super Cruise customers have logged more than 1.2 billion hands-free miles.

The real-world miles come from hundreds of GM data-collection vehicles and from retail customers who have opted in to share data. GM says it has optimized its training enough that an improvement spotted in the software can reach the road within a single day.

An OnStar AI assistant arrives on the 2027 Cadillac Lyriq via over-the-air update in the coming months. Anderson told Fortune a four-person team once had a voice prototype running in a Lyriq about 30 days after he asked for one.

Starting in 2027, a single OnStar app will replace OnStar Guardian and the separate Buick, Cadillac, Chevrolet and GMC apps, adding cameras that automatically record incidents; GM says it can’t access the recordings. GM Fleet, about 20% of annual sales, gets its own AI assistant, and OnStar Fleet has more than 2.4 million subscribers, which GM says is twice its nearest competitor.

None of this is cheap. GM says it is investing more than $16 billion in U.S. R&D and manufacturing this year, after about $7.6 billion in EV-related charges last year.