Public transcript

Anthropic AI Cuts Discounts for Best Customers - IPO is Failing

By: Eli the Computer Guy
Length: 21 min
Language: English
Text from: the video's captions
Transcribed: 4 October 2026

Read the full transcript Open on YouTube

Notes in English.

Summary

Eli the Computer Guy reacts to news that Anthropic ends discounts for enterprise customers once they hit their contractual spending caps, forcing renegotiation or higher prices. He compares it to Broadcom's punitive price hikes on VMware customers and argues it signals Anthropic's business model is so weak that even its best customers are a financial risk. He warns decision makers not to trust the AI frontier labs and to be cautious about deploying AI in production infrastructure.

Key points

Anthropic ends discounts once enterprise customers hit their contractual usage caps, forcing renegotiation or higher prices.
OpenAI, by contrast, gives customers the rest of the calendar month plus an extra month before reverting to list prices.
Amazon, Microsoft, and Google typically allow discounted pricing to continue beyond committed spending levels.
Eli compares the move to Broadcom's punitive VMware price hikes after the virtualization market saturated.
He argues AI valuations like Anthropic's $2 trillion target are disconnected from reality as free open source models erode the moat.
Palantir, Nvidia, and Booz Allen Hamilton are demanding new guarantees or reducing use of the labs' most advanced models.
The only logical explanation he sees is that Anthropic loses money on its best customers, meaning its business model is failing.
He advises being on the AI path but waiting until 'version three' before serious deployment.

Questions it answers

00:00Anthropic cuts discounts at caps

What is Anthropic doing to its best enterprise customers?

Anthropic is ending discounts once customers hit the usage limits in their contracts, requiring them to negotiate a new agreement or face higher prices. Sales teams contact customers saying they can avoid overages if they sign a new agreement immediately.

Anthropic ends discounts once customers hit contractual usage caps
Customers spending millions per year must renegotiate or face higher prices
OpenAI gives customers one to two months before reverting to list prices

00:43The VMware/Broadcom parallel

Why does he compare this to VMware and Broadcom?

After Broadcom bought VMware and the virtualization market was saturated, they punitively jacked up prices, figuring that even losing many customers they would make more money. He sees the same pattern of a company squeezing customers when growth is gone.

Broadcom jacked up VMware prices 4x to 6x after market saturation
Many lawsuits are ongoing over VMware pricing
Great technology can be ruined by a bad owning company

03:30Inflated AI valuations

Why are AI companies squeezing customers?

Companies like Anthropic, OpenAI, and SpaceX have valuations disconnected from reality, and they must prove their worth to investors as their moats degrade. He says this pressures them to start screwing customers in ways that degrade the customer experience.

SpaceX valued at $2.01 trillion, Anthropic seeking $2 trillion, OpenAI $1.5 trillion
Free open source models from Nvidia and Meta erode the frontier model moat
He argues the orchestration layer is more valuable than the models

06:54Weirdness of the discount policy

How does this discount behavior differ from industry norms?

Normally more usage earns bigger discounts, like buying in bulk at Costco, and Amazon, Microsoft, and Google let discounted pricing continue past committed spending. Anthropic instead penalizes customers the moment they hit their cap, which he finds bizarre.

Normally more usage means bigger discounts, like Costco
Anthropic's discounts are around 15% off list price
Over 100 firms spent more than $10 million each with Anthropic in 12 months

10:25Trust eroding in frontier labs

How are large customers responding to the frontier labs?

Firms doing sensitive corporate work like Palantir, Nvidia, and Booz Allen Hamilton are demanding new guarantees or reducing or eliminating use of the labs' most advanced models. He also cites Anthropic shutting off access to Mythos and Fable 5 within 90 minutes of a government notification as an unprecedented risk.

Palantir, Nvidia, and Booz Allen Hamilton are demanding guarantees or reducing use
Anthropic shut off Mythos and Fable 5 model access within 90 minutes of government notice
This is an unknown-unknown risk unlike anything in traditional IT

11:57Warning to decision makers

What should IT decision makers take away from this?

He warns that if Anthropic's business is so bad that its best customers are a risk, the AI bubble's wheels will come off badly. He advises being on the path of the AI stack but being very careful about deploying it, and waiting until 'version three' like with Microsoft products.

If Anthropic can't afford to serve its best customers, its business model is failing
Don't trust the frontier labs regardless of the technology quality
Be on the AI path but be very careful about production deployment

19:06Closing and announcements

What else does he mention at the end?

He announces he will restart free hands-on technology education in Baltimore, Maryland, with classes, interviews, and mentorship sessions under the Silicon Dojo vision, and points to the Meetup group link, Spotify, Apple Podcasts, and LinkedIn.

Free hands-on technology education returning in Baltimore, Maryland
Silicon Dojo vision continues with classes, interviews, and mentorship
Episodes available on Spotify and Apple Podcasts

Notes

The News

Anthropic is ending discounts for enterprise customers once they hit the usage caps in their contracts.
Customers spending millions or tens of millions of dollars per year must negotiate a new agreement or face higher prices.
Sales teams contact customers saying: "If we address this today with a new agreement, you can avoid overages."
Anthropic's discounts can be around 15% off listed price; more than 100 firms spent over $10 million each with Anthropic in the 12 months ending in June, and more than 1,000 firms spent over $1 million each.

The VMware/Broadcom Parallel

Eli compares this to Broadcom buying VMware (pre-2020 VMware was great) and then punitively jacking up prices once the virtualization market was saturated.
Broadcom's logic: quadruple the price and even if you lose many customers, you still make more money.
Many lawsuits are ongoing over VMware pricing.

Why Companies Squeeze Customers

AI frontier lab valuations (SpaceX at $2.01 trillion, Anthropic seeking $2 trillion, OpenAI at $1.5 trillion) have no connection to reality.
The moat of frontier models is degrading due to free open source models from Nvidia and Meta.
Eli argues the orchestration layer, not the frontier models, is the most valuable thing in AI right now.
To prove valuations to investors, companies start degrading the customer experience in legally permissible ways.

Industry Comparison

Amazon, Microsoft, and Google typically allow discounted pricing to continue for spending above the committed level.
OpenAI gives customers the rest of the calendar month plus an extra month to negotiate before reverting to list prices.
Normal utilization contracts reward more usage with bigger discounts, like buying in bulk at Costco.

Trust and Risk

Palantir Technologies, Nvidia, and Booz Allen Hamilton are demanding new guarantees or reducing/eliminating use of the labs' most advanced models.
Eli cites Anthropic shutting off access to Mythos and Fable 5 within 90 minutes of a government notification as an unknown-unknown risk.
The only logical explanation for the hard line is that Anthropic loses money on these customers, meaning its business model is so bad its best customers are a risk.

Advice

Don't trust the frontier labs; business isn't personal.
Be on the path of the AI stack but be very careful about deploying it — think "wait till version three."
He plans to restart free hands-on technology education in Baltimore, Maryland (Silicon Dojo).

Text from the video's own captions. Summary and notes written by AI from the transcript, so check anything important against the recording.