Nvidia’s ‘AI Factory’ narrative faces reality check at Transform 2025

Gloves came out Tuesday in VB Transform 2025 Alternative chip producers protested directly in a panel during the prevailing narrative of NVIDIA in a panel: AI result may Does a “factory” and command command 70% of the total edges?

Jonathan Ross, CEO GleHe did not justify the words when discussing NVIDIA’s carefully crafted messaging. “AI factory is a marketing way to scare the sound of a sound to ai,” said Ross during the panel said. Sean Lies, CTO BrainAn opponent was equally directly directed directly: “I do not think that fighting for every penny of all service providers who are all over there, sitting comfortably with 70 points.”

Hundreds of billions of infrastructure investment and enterprise AI enterprise are in danger of future architecture. Currently, it is currently locked in weekly negotiations with open and other providers for the AI ​​leaders, Panel, AI initiatives announced concerned about why they continue to hit road obstacles.

>> – siSee all transforms 2025 coverage of our entire transform<

No one speaks about potential crisis

“Anyone in fact, which is a great user from this Gen AI, knows that you can go to Openai or someone, and they will really be able to serve you enough verses,” said Dylan Patel, founder Semicolar. There are weekly meetings to try to convince some of the largest AI users and their model providers. Then there are weekly meetings between these model providers and their device providers. “

Panel participants also pointed to the significance as the significance exposes the fundamental defect in the factory. Traditional production, adds the required signals by adding capacity. However, when the enterprises require 10 times more effective, they discover that the supply chain is twisted. GPUS requires two years of lead time. Information centers need permits and power contracts. Infrastructure has not been built for exponential scale, providers force providers to enter the ration with API limits.

According to the Patel, Anthropical In total, $ 2 billion jumped for $ 3 billion in six months. Cursor essentially gone from scratch to $ 500 million. Open Exceeded $ 10 billion. Still enterprises still can’t get the verses needed.

Why ‘factory’ thinks the AI ​​economy

Jensen Huang’s “AI factory“Concept, expenditures in reduces standardization, commodity and efficiency. However, the panel revealed three fundamental paths of this metaphor:

First, infertility is not uniform. “Even today, say, DeepSseek, Deepseek, there are a number of providers along the curve of which they provide how fast,” said Patel. DeepSseek serves its model at the lowest price, but provides 20 tokens per second. “No one wants to use a model in 20 verses in one second. I speak faster than 20 points.”

Second, the quality changes wildly. Ross pulled a parallel in the standard oil history: “When the standard oil begins, when the oil has different quality. You can get fat from a seller and turn your home.” The results of today’s AI are facing various methods to reduce the cost of damaging the quality of fruitless compromise, with similar quality change.

The third and the most critical, economy turns upside down. “One of the things that is unusual from the EU is that you can’t spend more to get better results,” Ross said. “You can’t simply apply a program, say, I will spend twice a lot to host my program and improve applications.”

Ross noted that Mark Zuckerberg said, when “the only ones that start full quality”, I explained the quality crisis of the industry. It was not just recognized. This was the indictment of cutting corners of each provider.

Ross wrote the mechanic: “Many people, a lot of tricks to reduce quality, but reduce the costs, do much to improve their speeds.” Techniques are technically voiced, but the effect is true. Reduces quantity accuracy. Puts the pruning settings. Each optimization worsens the model performance in businesses, and production cannot be found until the production fails.

Standard oil covers Parallel Ross Drew shares. Today’s result market faces the problem of the same quality. Providers who bet on enterprises do not matter between 95% and have the subtleties to measure convictions against companies such as 100% accuracy meta.

This creates immediate imperatives for enterprise buyers.

1. Create quality criteria before choosing providers.
2. Check your existing account partners for unspecified optimization.
3. Accept this award price for absolutely loyalty, now a permanent market feature. When Zuckerberg’s difference is the difference, the period of assessing the functional equivalence among results ended.

1 Million dollar Token Paradox

The most obvious moment came when the panel discussed the price. Stressed a concern for false industry.

This observation cuts the heart of the discovery problem of AI’s price. These verses compete to drive token costs below $ 1.50 per million when changing each side of the case. The panel has openly agreed with each other where the mathematics has not been added.

“Everyone spends all the growing beginnings, and the amount they spend in a service for verses is almost compatible with each other,” he said. Compared to the label of this 1: 1 to the AI ​​verses, the panel participants ignored the “Factory” narratio.

Performance changes everything

Cerebras and Grog are not only competitive; They also compete on performance. They radically change what is possible in terms of input. “With the tech technology that we set up, today, today, sometimes 50 times, more than 50 times, faster performance,” we are lying, “he said.

This is not a growing development. Allows entirely new use. “We have clients with agency work flows that can take 40 minutes and want this work to work in real time,” he said. “Though these things are ready to pay only the top dollar, it is not possible.”

The speed of speed creates a bifurcated market that prevents the factory standardization. The same infrastructure in need of real-time results for the customer facing applications cannot use the same infrastructure as those who work in the bulk processes.

Real Bottleneck: Power and Information Centers

While everyone pays attention to the supply of the chip, the panel revealed the real restriction of the AI ​​deployment. “The power of the data center is a big problem. You really do not find the data center space in the United States,” said Patel. “Power is a big problem.”

The infrastructure calls goes beyond the production of fundamental resource restrictions. As Patel explains, “TSMC is $ 200 million in Taiwan, isn’t it?

However, the production of the chip means nothing without infrastructure. “The reason for us to see these great Middle Eastern deals and both of these companies have great jobs in the Middle East, this is the power.” Global scramble for calculation, electricity that can build these electrical systems, where electricity is caught in anywhere, where it is going to go to the world to start. “

Google’s ‘Success Disaster’ Everyone becomes a reality

Ross shared an anecdote from Google’s history: “In 2015 there was a period of very popular in Google.

This example is now repeated in placing each entity in the AI. Applications or traction or hockey stick growth cannot immediately learn the infrastructure limits. There is no middle ground, there is no smooth-scale curve where the plant economy will predict.

What does this enterprise mean for the AI ​​strategy

For CIOS, CISOS and AI leaders, the panel’s verses require strategic revaluation:

Capacity building requires new models. Traditional this prediction line takes growth. AI workload breaks this hypothesis. When successful applications increase the average monthly token consumption, annual potential pedestrians are obsolete in the quarter. Enterprises must change from static procurement periods for dynamically capacity management. Set contracts with Burst provisions. Use weekly, not quarterly. Accept this AI-scale patterns, it is like viral adoption curves that do not have a traditional enterprise program.

The speed rewards are permanent. The idea of ​​issuing the result at a single price, ignores the massive performance gaps between providers. The businesses need a budget from where it is more important.

Beats architectural optimization. Grog and Cerebras do not win by making GPU better. AI calculation wins the basic architecture again. Enterprises who bet on everything in GPU-based infrastructure can be stuck in a slow lane.

Power infrastructure is strategic. The restriction chips or program, but not a kilowatt and cooling. Smart enterprises are already locked in the field of power capacity and data center in 2026 and outside.

The infrastructure cannot ignore reality institutions

The panel revealed a fundamental fact: AI factory metaphor is not only wrong, but also dangerous. Planning for a market that does not exist in strategies and standardized delivery of businesses around the commodity result.

The actual market is working on three brutal truths.

1. Suppliers create force inversions that are begging for the deductions of terms and enterprises.
2. The difference between quality contradictions, 95% and 100% accuracy determines whether AI applications are successful.
3. Set the binding restrictions in infrastructure restrictions, not technology, but aI transformation.

The road forward for Cisos and AI leaders requires completely to give up the factory. Now lock in energy capability. Inspection time providers for hidden quality degradation. Set up seller connections based on architectural advantages, not the savings of marginal cost. Most critically, accept 70% margin’s payment for a reliable, high-quality result, can be the most intelligent investment.

Alternative chip manufacturers in Transform did not try the NVIDIA narrative. The enterprises have shown that they are facing an option: pay for quality and performance or join the weekly negotiating meetings. The Convention of the panel was clear: success, a dimensional compliance – requires the adaptation of concrete workloads in accordance with the relevant infrastructure, rather than continuing all solutions.