Everyone online talks about EVs dominating drag races, but as it turns out, the fastest 0-60 production car is actually a gas car: the Dodge Demon 170 runs roughly 1.7 seconds on a prepped surface. 0-60 became the EV’s signature metric, just like phone benchmark scores became the signature metric for chips. The question is, why 0-60 specifically — not Nürburgring lap times, not 100–200 km/h acceleration, not ten consecutive laps without power fade?
Because 0-60 is the metric EVs win most easily. Low-speed launches primarily depend on tire grip, all-wheel-drive torque distribution, and instantaneous torque response. EV motors deliver full torque from zero RPM, and all-wheel-drive control can be tuned entirely through software. So even a two-and-a-half-ton electric sedan can clock a sub-2-second 0-60. This metric requires no driving skill, no track experience, no holistic understanding of vehicle dynamics. It only needs a sufficiently large motor and a sufficiently clever launch control.
But 0-60 gets trotted out repeatedly not just because it’s easy to win, but because it’s the most suitable for cyber cockfighting. Most people lack the ability to comprehensively evaluate a car’s strengths and weaknesses. Lap times depend on track conditions, driving feedback depends on personal feel, resale value depends on time — these are all too subjective, too complex, too hard to argue over. 0-60 is different: a single number, higher or lower, instantly clear. Fans don’t need to understand chassis engineering or have track experience; they just need the number from their preferred brand to be bigger to declare victory. This is the essence of cyber cockfighting: pick an objective metric that everyone can understand, then fight over it. What’s being contested is fan identity, not product judgment.
Cyber cockfighting can determine whose supporters are louder, but it can’t settle a more useful question: what is this thing actually worth. 0-60 can tell you whether the powertrain is strong, but a car’s price has never depended solely on a single number.
The way high-performance EVs are built follows the same playbook as American muscle cars. Big battery, high-power motor, big brakes, wide tires — all stacked around an absurd power figure. This is exactly the same logic behind the Hellcat and Demon: cram in a large-displacement supercharged V8 first, then build the drivetrain, suspension, tires, and cooling around it. The technical medium has changed; the spirit hasn’t: solve problems with overwhelming power.
The earlier article on chassis discussed the “brute force works wonders” engineering philosophy of EVs: the battery is too heavy, the strategic layer is locked in, so the only option is an arms race at the tactical level. That judgment holds at the chassis level, and it holds at the entire product definition level too. EVs aren’t designed as sports cars from the ground up — they’re given an outrageous horsepower number first, then tamed with big brakes, wide tires, aggressive cooling, and software control.
Muscle cars were never expensive. The Hellcat starts at $70,000, and the Demon 170 tops out around $200,000. This price band is no accident. A single-point extreme product has a naturally low pricing ceiling, because you sacrifice every other dimension to feed that one metric. Weight goes up, thermal management becomes difficult, brake feel degrades, driving feedback disappears, everyday usability is compromised. What the buyer gets is a report card, not a complete car.
The race track is the proving ground for the single-point extreme approach. Porsche 911 GT3, 502 horsepower, Nürburgring in 6:59. Yangwang U9 Xtreme, over 3,000 horsepower, Nürburgring in 6:59. Six times the power, same lap time.
The U9 Xtreme is not the standard U9. BYD reworked the cooling system, swapped in carbon-ceramic brakes, mounted GitiSport semi-slick tires, and limited production to 30 units. Clocking 6:59 is already a hard-won achievement for EV engineering. But what this result demonstrates is not “EVs outperform gas cars” — it’s that “brute force works wonders” has an efficiency ceiling on the track. The GT3 achieves the same lap time with 502 horsepower through the matrix efficiency of weight, aerodynamics, chassis geometry, braking consistency, and thermal management. Every single horsepower is pushed to its limit, rather than relying on an extra two thousand horsepower to brute-force physics.
Car and Driver’s Lightning Lap data from Virginia International Raceway tells the same story. The Corvette Z06 laps in 2:38, faster than the Ferrari 296 GTB and Lamborghini Huracán Performante, starting at $110,000. The Ferrari 296 GTB starts at over £250,000 in the UK. If faster lap times justified higher prices, the Corvette should cost the most. But there is no linear relationship between lap times and price.
In July 2026, Ferrari launched the 12Cilindri Manuale: an 830-horsepower naturally aspirated V12, six-speed manual transmission, limited to 1,499 units, starting at €590,000. The standard 12Cilindri uses an automatic dual-clutch and starts at €395,000. The manual version costs 50% more than the automatic and has a lower top speed: 315 km/h in manual mode versus 340 km/h in automatic mode. Slower, more expensive, harder to drive. This makes no sense in a “quantifiable” framework. The manual isn’t as fast as the automatic, shifts have delay, the clutch has a learning curve, it’s exhausting in traffic. But in Ferrari’s framework, it makes perfect sense: a manual transmission is irreplaceable. An automatic dual-clutch can be sourced by any manufacturer, tuned by any engineer. A manual V12 with a gated shifter and a clutch pedal is a concrete artifact of Ferrari’s seventy-year narrative. It’s not selling performance — it’s selling something you can only buy here.
Lamborghini took a different path. It never raced in F1, and its lap times were never best-in-class. But the naturally aspirated V12 engine, scissor doors, and low, wide, exaggerated stance are social currency the moment it’s parked. It doesn’t sell track results — it sells drama and symbolic value. When you buy a Lambo, what you’re buying isn’t a 0.2-second faster 0-60. It’s the very fact of “I drive a Lambo.”
Quantifiable strength will inevitably be caught up. The muscle car’s 0-60 was caught by EVs, and the EV’s lap time will be caught by the next high-power EV. Once a dimension can be quantified and compared, competition converges to cost, and pricing power goes to zero.
Expensive things are expensive because they are irreplaceable. Ferrari’s irreplaceability comes from seventy years of F1 narrative accumulation, limited allocation strategy, and client vetting — you don’t just pay to get one, you have to be a “qualified Ferrari client.” An LV bag is worse at carrying things than a woven plastic sack, yet its pricing power dwarfs the sack’s, because a century of craftsmanship narrative and brand symbolism are irreplaceable. Maybach’s irreplaceability comes from the very posture of “we don’t compete on specs”: it doesn’t talk about horsepower, doesn’t talk about lap times — it tells the image of a man in a suit stepping off a private jet into the back seat.
The industrial product logic goes: quantifiable → replicable → competition converges → cost → pricing power zero. The artisanal product logic goes: unquantifiable → irreplaceable → competition doesn’t converge → pricing power exists. Quantifiability is the ticket to entry, not pricing power. You need a passable 0-60, a passable lap time, a passable benchmark score just to earn a seat at the table. But once you’re at the table, what determines whether you can command a premium is the things that can’t be replicated.
AI benchmark culture is 0-60 culture. Whichever model scores highest on MMLU, dominates SWE-bench, leads on HumanEval — that gets called “the strongest AI.” This is no different from car fans using 0-60 for cyber cockfighting: pick a number, whoever’s number is bigger wins, no engineering understanding required, no need to look at real-world performance, fans declare victory on the spot.
When technical people face the difficulty of evaluating AI, their first instinct is to top the leaderboard — and this is entirely understandable. Users and investors need anchors. If you say “my model feels better to use” but can’t produce numbers, they won’t even put you on the shortlist. Ferrari can afford not to compare lap times because it has seventy years of racing narrative and scarcity built up. But a new brand without that accumulation — users’ first reaction is “why should you even be on the shortlist.” So benchmark performance is the ticket to entry: prove you’re not a toy first, then earn the chance to sit down and talk.
But the moment you top the leaderboard, you put yourself on the price-comparison shelf. You say your MMLU is 91, and procurement immediately asks: another model scores 89 but is 40% cheaper, why should I buy yours? This is the same logic behind muscle cars never selling at a premium. No matter how fast the Hellcat’s 0-60, it can’t command a Ferrari price, because it placed itself on a table of horsepower, acceleration, and price — and there will always be a cheaper competitor on that table. When you choose to compete on a quantifiable dimension, competition converges to cost-performance ratio.
More troubling is that benchmarks are public test sets. Models can be optimized specifically for them, or even have test questions mixed into training data to inflate scores. Once this is widely known, scores degrade from evidence to clues. The car world already scrutinizes 0-60 by asking about tires, surface, rollout, altitude; AI users will likewise ask about training set contamination, prompt templates, cherry-picking. At that point, a single high score triggers due diligence instead of trust. What truly matters is not the number on the leaderboard — it’s the eval users can rerun on their own data, the transparency around failure cases, whether the product actually works well once embedded in their real workflow.
An MIT study tracked model usage on OpenRouter over five months and laid this dilemma bare in numbers. Open-source models reach 90% of closed-source model performance at one-sixth the cost, and catch up to closed-source benchmark scores within 13 weeks of release. Yet closed-source models still account for 80% of usage and 96% of revenue. The benchmark gap is already tiny, but pricing power remains with the closed-source models. It’s the same as cars: the 0-60 gap is already tiny — Demon 170 at roughly 1.7 seconds, Plaid at roughly 1.9 seconds, GT3 at roughly 2.9 seconds — but pricing power doesn’t belong to the car with the fastest 0-60.
Benchmarks can explain where attention comes from; they can’t explain why money stays. They are a traffic entry point, not a balance sheet.
That 13-week figure from the MIT study has another layer of meaning: if you’re an AI builder working on closed-source models, your benchmark advantage has a shelf life of just 13 weeks. Open source will catch up to your scores, but it won’t catch up to the customer data you onboarded, the permission systems you tuned, the evals you accumulated, the workflows you embedded during those 13 weeks. Continuing to chase the next leaderboard will keep you stuck on the track forever, with each lead only lasting one quarter. Those 13 weeks are construction season, not harvest season. Turning a high score into a default integration position is what allows you to keep charging after the score advantage goes to zero.
The most dangerous strategy is neither topping the leaderboard nor refusing to — it’s having an advantage before purchase but zero stickiness after. Users come drawn by leaderboard rankings, try the product, find it can’t handle real workflows, and churn within two weeks. Exactly like a muscle car: jaw-dropping 0-60 in the showroom, but you drive it home and realize it’s not a complete car. What’s truly fatal is a split between the acquisition story and the retention story: the former runs on scores, while the latter has no irreplaceable usage inertia whatsoever.
So how do you actually sell what benchmarks can’t measure? AI products with real pricing power compete on things like the quality of context infrastructure, the reliability of feedback loops, the depth of workflow embedding, and the accumulation of user trust. These things can’t go on a spec sheet, can’t be countered by a competitor saying “we’ll match and exceed,” and that’s precisely why they can anchor price. But users won’t pay just because you write “better context management.” The unquantifiable must first be made experiential. A user imports a real project, and the AI points out three constraints only a veteran colleague would know — that aha moment delivers exactly the unquantifiable value of “good context infrastructure.” The solution to cold start isn’t explaining capability; it’s designing a differentiated experience that is guaranteed to happen within 20 minutes.
AI products have their own “matrix” too. A good AI product is more than just the model itself — it includes context retrieval, tool calling, permission boundaries, latency, cost, observability, error correction mechanisms, UI workflow, and team collaboration. The efficiency with which these dimensions work together is what determines whether the product is actually good. A mid-tier model paired with an excellent repo map, task decomposition, test feedback, and rollback-capable execution can outperform a bare flagship model in real output. The GT3’s speed doesn’t come from horsepower — it comes from brakes, chassis, tires, and aerodynamics working in concert. An AI product’s speed doesn’t come from benchmark scores either — it comes from the collaborative efficiency of the entire matrix.
Thirteen weeks later when open source catches up, if all you have left is an outdated screenshot, benchmark parity will directly destroy your pricing. But if users have already migrated their workflows in during those 13 weeks, what benchmark parity changes is only your underlying model procurement cost, not your customer relationships. A model that tops benchmarks but is unreliable in actual use, and a car that does 1.7-second 0-60 but can’t beat a 500-horsepower GT3 on track — they are the same thing. They are both single-point extreme, system-level suboptimal solutions. Their scores mean something — but scores are not pricing power.