Be a part of the occasion trusted by enterprise leaders for practically twenty years. VB Rework brings collectively the folks constructing actual enterprise AI technique. Be taught extra
The gloves got here off at Tuesday at VB Rework 2025 as various chip makers instantly challenged Nvidia’s dominance narrative throughout a panel about inference, exposing a basic contradiction: How can AI inference be a commoditized “manufacturing unit” and command 70% gross margins?
Jonathan Ross, CEO of Groq, didn’t mince phrases when discussing Nvidia’s fastidiously crafted messaging. “AI manufacturing unit is only a advertising and marketing technique to make AI sound much less scary,” Ross mentioned in the course of the panel. Sean Lie, CTO of Cerebras, a competitor, was equally direct: “I don’t assume Nvidia minds having all the service suppliers preventing it out for each final penny whereas they’re sitting there snug with 70 factors.”
Tons of of billions in infrastructure funding and the long run structure of enterprise AI are at stake. For CISOs and AI leaders presently locked in weekly negotiations with OpenAI and different suppliers for extra capability, the panel uncovered uncomfortable truths about why their AI initiatives maintain hitting roadblocks.
>>See all our Rework 2025 protection right here<<The capability disaster nobody talks about
“Anybody who’s truly a giant person of those gen AI fashions is aware of which you can go to OpenAI, or whoever it’s, they usually gained’t truly have the ability to serve you adequate tokens,” defined Dylan Patel, founding father of SemiAnalysis. There are weekly conferences between among the largest AI customers and their mannequin suppliers to attempt to persuade them to allocate extra capability. Then there’s weekly conferences between these mannequin suppliers and their {hardware} suppliers.”
Panel individuals additionally pointed to the token scarcity as exposing a basic flaw within the manufacturing unit analogy. Conventional manufacturing responds to demand alerts by including capability. Nonetheless, when enterprises require 10 instances extra inference capability, they uncover that the availability chain can’t flex. GPUs require two-year lead instances. Knowledge facilities want permits and energy agreements. The infrastructure wasn’t constructed for exponential scaling, forcing suppliers to ration entry by API limits.
In line with Patel, Anthropic jumped from $2 billion to $3 billion in ARR in simply six months. Cursor went from basically zero to $500 million ARR. OpenAI crossed $10 billion. But enterprises nonetheless can’t get the tokens they want.
Why ‘Manufacturing unit’ considering breaks AI economics
Jensen Huang’s “AI manufacturing unit” idea implies standardization, commoditization and effectivity positive factors that drive down prices. However the panel revealed three basic methods this metaphor breaks down:
First, inference isn’t uniform. “Even right this moment, for inference of, say, DeepSeek, there’s a lot of suppliers alongside the curve of type of how briskly they supply at what value,” Patel famous. DeepSeek serves its personal mannequin on the lowest value however solely delivers 20 tokens per second. “No one needs to make use of a mannequin at 20 tokens a second. I speak quicker than 20 tokens a second.”
Second, high quality varies wildly. Ross drew a historic parallel to Customary Oil: “When Customary Oil began, oil had various high quality. You possibly can purchase oil from one vendor and it’d set your own home on hearth.” As we speak’s AI inference market faces comparable high quality variations, with suppliers utilizing numerous strategies to cut back prices that inadvertently compromise output high quality.
Third, and most critically, the economics are inverted. “One of many issues that’s uncommon about AI is which you can’t spend extra to get higher outcomes,” Ross defined. “You may’t simply have a software program software, say, I’m going to spend twice as a lot to host my software program, and functions can get higher.”
When Ross talked about that Mark Zuckerberg praised Groq for being “the one ones who launched it with the total high quality,” he inadvertently revealed the trade’s high quality disaster. This wasn’t simply recognition. It was an indictment of each different supplier reducing corners.
Ross spelled out the mechanics: “Lots of people do loads of methods to cut back the standard, not deliberately, however to decrease their value, enhance their velocity.” The strategies sound technical, however the affect is easy. Quantization reduces precision. Pruning removes parameters. Every optimization degrades mannequin efficiency in methods enterprises could not detect till manufacturing fails.
The Customary Oil parallel Ross drew illuminates the stakes. As we speak’s inference market faces the identical high quality variance downside. Suppliers betting that enterprises gained’t discover the distinction between 95% and 100% accuracy are betting in opposition to corporations like Meta which have the sophistication to measure degradation.
This creates instant imperatives for enterprise consumers.
- Set up high quality benchmarks earlier than choosing suppliers.
- Audit present inference companions for undisclosed optimizations.
- Settle for that premium pricing for full mannequin constancy is now a everlasting market function. The period of assuming purposeful equivalence throughout inference suppliers ended when Zuckerberg known as out the distinction.
The $1 million token paradox
Essentially the most revealing second got here when the panel mentioned pricing. Lie highlighted an uncomfortable reality for the trade: “If these million tokens are as worthwhile as we consider they are often, proper? That’s not about shifting phrases. You don’t cost $1 for shifting phrases. I pay my lawyer $800 for an hour to jot down a two-page memo.”
This commentary cuts to the guts of AI’s worth discovery downside. The trade is racing to drive token prices beneath $1.50 per million whereas claiming these tokens will remodel each side of enterprise. The panel implicitly agreed with one another that the mathematics doesn’t add up.
“Just about everyone seems to be spending, like all of those fast-growing startups, the quantity that they’re spending on tokens as a service virtually matches their income one to 1,” Ross revealed. This 1:1 spend ratio on AI tokens versus income represents an unsustainable enterprise mannequin that panel individuals contend the “manufacturing unit” narrative conveniently ignores.
Efficiency adjustments every thing
Cerebras and Groq aren’t simply competing on worth; they’re additionally competing on efficiency. They’re basically altering what is feasible when it comes to inference velocity. “With the wafer scale know-how that we’ve constructed, we’re enabling 10 instances, generally 50 instances, quicker efficiency than even the quickest GPUs right this moment,” Lie mentioned.
This isn’t an incremental enchancment. It’s enabling fully new use instances. “Now we have clients who’ve agentic workflows which may take 40 minutes, they usually need these items to run in actual time,” Lie defined. “These items simply aren’t even potential, even in the event you’re keen to pay high greenback.”
The velocity differential creates a bifurcated market that defies manufacturing unit standardization. Enterprises needing real-time inference for customer-facing functions can’t use the identical infrastructure as these operating in a single day batch processes.
The actual bottleneck: energy and information facilities
Whereas everybody focuses on chip provide, the panel revealed the precise constraint throttling AI deployment. “Knowledge middle capability is a giant downside. You may’t actually discover information middle area within the U.S.,” Patel mentioned. “Energy is a giant downside.”
The infrastructure problem goes past chip manufacturing to basic useful resource constraints. As Patel defined, “TSMC in Taiwan is ready to make over $200 million value of chips, proper? It’s not even… it’s the velocity at which they scale up is ridiculous.”
However chip manufacturing means nothing with out infrastructure. “The rationale we see these massive Center East offers, and partially why each of those corporations have massive presences within the Center East is, it’s energy,” Patel revealed. The worldwide scramble for compute has enterprises “going internationally to get wherever energy does exist, wherever information middle capability exists, wherever there are electricians who can construct these electrical techniques.”
Google’s ‘success catastrophe’ turns into everybody’s actuality
Ross shared a telling anecdote from Google’s historical past: “There was a time period that turned extremely popular at Google in 2015 known as Success Catastrophe. A number of the groups had constructed AI functions that started to work higher than human beings for the primary time, and the demand for compute was so excessive, they had been going to want to double or triple the worldwide information middle footprint rapidly.”
This sample now repeats throughout each enterprise AI deployment. Purposes both fail to achieve traction or expertise hockey stick development that instantly hits infrastructure limits. There’s no center floor, no clean scaling curve that manufacturing unit economics would predict.
What this implies for enterprise AI technique
For CIOs, CISOs and AI leaders, the panel’s revelations demand strategic recalibration:
Capability planning requires new fashions. Conventional IT forecasting assumes linear development. AI workloads break this assumption. When profitable functions improve token consumption by 30% month-to-month, annual capability plans change into out of date inside quarters. Enterprises should shift from static procurement cycles to dynamic capability administration. Construct contracts with burst provisions. Monitor utilization weekly, not quarterly. Settle for that AI scaling patterns resemble these of viral adoption curves, not conventional enterprise software program rollouts.
Velocity premiums are everlasting. The concept inference will commoditize to uniform pricing ignores the large efficiency gaps between suppliers. Enterprises must finances for velocity the place it issues.
Structure beats optimization. Groq and Cerebras aren’t successful by doing GPUs higher. They’re successful by rethinking the basic structure of AI compute. Enterprises that guess every thing on GPU-based infrastructure could discover themselves caught within the gradual lane.
Energy infrastructure is strategic. The constraint isn’t chips or software program however kilowatts and cooling. Sensible enterprises are already locking in energy capability and information middle area for 2026 and past.
The infrastructure actuality enterprises can’t ignore
The panel revealed a basic reality: the AI manufacturing unit metaphor isn’t solely unsuitable, but in addition harmful. Enterprises constructing methods round commodity inference pricing and standardized supply are planning for a market that doesn’t exist.
The actual market operates on three brutal realities.
- Capability shortage creates energy inversions, the place suppliers dictate phrases and enterprises beg for allocations.
- High quality variance, the distinction between 95% and 100% accuracy, determines whether or not your AI functions succeed or catastrophically fail.
- Infrastructure constraints, not know-how, set the binding limits on AI transformation.
The trail ahead for CISOs and AI leaders requires abandoning manufacturing unit considering fully. Lock in energy capability now. Audit inference suppliers for hidden high quality degradation. Construct vendor relationships primarily based on architectural benefits, not marginal value financial savings. Most critically, settle for that paying 70% margins for dependable, high-quality inference could also be your smartest funding.
The choice chip makers at Rework didn’t simply problem Nvidia’s narrative. They revealed that enterprises face a alternative: pay for high quality and efficiency, or be part of the weekly negotiation conferences. The panel’s consensus was clear: success requires matching particular workloads to applicable infrastructure moderately than pursuing one-size-fits-all options.
