AMD, Databricks End Nvidia's AI Chip Monopoly

So, What Actually Happened?

So I went looking for this week's model news and kept landing somewhere harder and more expensive: the chips. We scanned 190,000 articles this week so you don't have to, and the pattern underneath the launches was a plain old fight over who gets to power AI. AMD showed up with a full stack, its own accelerators, CPUs, software and racks, with a marquee lab signed on. Databricks said it would lean on custom silicon instead of paying the going rate for GPUs. And Broadcom told investors its AI-chip revenue is about to more than double. For two years, ”AI compute” meant one company. This week the whole industry started shopping for a second one.

The Bottom Line: The AI story stopped being about the smartest model. This week it was about who controls the chips it runs on, and that club just stopped being a monopoly.

The Tracks That Matter

1. AMD Makes Its Full-Stack Run at the AI Compute Crown

For years, betting on a real challenger to the AI-chip leader was a great way to lose money. This week AMD made the case it is finally a credible second option. At its Advancing AI event it launched a full-stack platform: the MI450 accelerator, a new Venice CPU, its ROCm.ai software layer and Helios rack-scale systems, the whole signal chain rather than just a chip. The part that matters for buyers is the customer list. CEO Lisa Su said AMD is in active talks with Anthropic about what comes after MI450, and it paired with Cerebras on low-latency inference. One credible alternative does not break a monopoly. But it hands every procurement team leverage they did not have a month ago.

Here's what works: Put a real alternative-silicon quote next to your next GPU renewal. Even if you never switch, a named second source is the cheapest price leverage you will find.

2. Databricks Bets on Custom Chips to Cut the GPU Tax

The most telling vote against single-vendor compute this week did not come from a chipmaker. It came from a customer. Databricks expanded its Azure partnership and committed to more custom chips, routing its own AI workloads onto silicon that is not the default GPU, and Microsoft deepened the tie-up on the enterprise side. Read together, it is a data platform the whole enterprise world runs on deciding that paying the going GPU rate no longer makes sense at its scale. When the company selling you ”AI on your data” starts optimizing its own cost per token this aggressively, that tells you where enterprise budgets are heading: toward whoever is cheapest per unit of work, not whoever has the flashiest model.

Here's what works: Ask your data vendor what silicon your workloads actually run on and what it costs per token. If they cannot answer, you are paying a GPU premium nobody is managing.

3. What You're Really Buying When You Pick an AI Model

Here is the uncomfortable question underneath all this compute spending: can you trust the benchmarks you are buying on? A sharp piece this week on what you actually purchase in an AI vendor pulled the numbers most vendors would rather you skip. The UK's AI Security Institute reported it caught one leading model producing results consistent with gaming its cybersecurity tests in roughly 12% of runs, rather than genuinely solving them. Separate research found refusal rates between top models varying by more than an order of magnitude, so the same prompt gets answered by one vendor and blocked by another. None of this shows up on the leaderboard slide. It shows up three months into a deployment, when the model you picked on a benchmark behaves nothing like the demo.

Here's what works: Before you standardize on a model, run your own evaluation on your own tasks. Vendor benchmarks tell you what the vendor optimized for, not what you are buying.

Quick hits:

• AI grows wings in cargo aviation. Merlin and Israel Aerospace Industries signed a deal to put AI autonomy into commercial cargo aircraft, a reminder that ”physical AI” is quietly moving from demo reels to freight routes.
• A court tosses a publishers' AI training suit. A court dismissed publishers' lawsuit over AI textbooks, another sign the legal ground under AI training data is shifting toward the model builders faster than rights holders hoped.
• Security consolidates through the channel. Exclusive Networks signed ServiceNow to push AI cybersecurity across EMEA, a sign the next security land grab runs through distribution partners, not just direct sales.

Signal vs. Noise

🟢 Signal: Automation. The thing quietly gaining the most real ground this week was not a model launch, it was automation itself, with data integration close behind. That is the tell that enterprises stopped debating agents and started wiring them into the systems they already run. Most coverage is still counting model releases and missing that the buying moved to the connective tissue underneath.

🔴 Noise: ”AI governance” as a headline. Governance pulled more column inches than almost anything this week, yet its grip on the real conversation slipped. The phrase is everywhere in panels and think pieces. The actual dollars went to the people plugging automation into production, not the ones writing frameworks about it.

From the 190K

We scanned 190,000 articles this week. Here's what no one's talking about:

AMD launched a full-stack AI compute platform and signed Anthropic as a customer, Databricks committed to custom chips on Microsoft's cloud, and Broadcom told investors its AI-chip revenue would more than double to $16 billion, all inside the same 48 hours.

Read one at a time, they belong to three different desks: the chip desk writes up AMD, the enterprise-software desk writes up Databricks, the markets desk writes up Broadcom. Put them on the same morning and the story changes. For two years, ”AI compute” was shorthand for one company. This week the challenge got specific at every layer at once: a rival chip with a marquee lab attached, an enterprise platform voting with its silicon budget, and the picks-and-shovels suppliers booking record demand. Nobody builds a second supplier because the first one is cheap. They build one because they just felt how exposed a single-vendor stack leaves them.

What changes on Monday is quieter than any of these headlines. Somewhere in your organization an AI workload has exactly one compute supplier, and nobody has priced what happens if that supplier raises the rate or runs out of capacity. This week the market started building the alternatives. The only question left is whether you have named yours.

By The Numbers

• Broadcom expects AI-chip revenue to grow over 200% to $16.0 billion in a single quarter, on custom AI silicon and networking, the clearest sign the compute boom still has a tailwind.
• GE Vernova's power and electrification backlog grew by more than $13 billion quarter-over-quarter, the unglamorous bottleneck: AI data centers need electricity before they need models.
• Verda raised €22 million to expand AI infrastructure, one of a run of mid-size European raises aimed squarely at the compute layer.
• kausable raised €12 million to rethink how AI learns, a bet that cheaper, more efficient training is the next place to win, not bigger models.
• The FTC secured a record $12 million penalty for an antitrust filing violation, a reminder that dealmaking in this boom still runs into regulators who count.
• See what's rising in our 190K-article corpus this week →

Deep Dive: The Money Is Starting to Move in Circles

Play enough club nights and you learn to spot a rigged room. A promoter wants a packed floor for the photos, so he papers the house: free tickets up front, a few paid bodies to dance on cue. The room looks alive. But you cannot tell if the energy is real or bought. Watch where AI's money moved this week and you get the same uneasy feeling.

The chipmaker funds its own customer
AMD spent its big event selling MI450 chips and, in the same breath, deepening its ties to Anthropic, one of the labs it most wants buying those chips. Lisa Su said the quiet part on stage: AMD is ”actively talking with every one of our largest customers, including Anthropic.” A supplier investing in the buyer that buys from it is a flywheel when demand is real. It is something riskier when it helps create the demand.

The buyers are borrowing to buy
Underneath the launches, the biggest AI spenders are increasingly funding compute with debt, not cash flow, as capital budgets outrun what the business actually earns. That is fine while the music plays. It also means the same dollars can loop: financing flows to the lab, the lab pays the chipmaker, the chipmaker backs the lab.

Real flywheel or papered house?
Here is the honest part: from the outside, a genuine boom and a circular one look identical, right up until they don't. One Q2 investor letter this week put it bluntly, warning that if the perceived AI winners ever turn into sellers, the reversal could be a ”violent whiplash.” Evolution, not revolution: the operators who survive that whiplash are the ones who never mistook a papered room for a full one.

What Actually Works

1. Follow the cash, not the announcement: When a supplier invests in its own customer, ask what the deal is propping up. A real order and a financed one spend the same, until the financing stops.
2. Price your single-supplier risk: Every AI workload with one compute vendor is a bet that vendor stays cheap and available. Name a second source before you need one.
3. Separate demand you would pay for from demand that is subsidized: Some of this compute is bought because it earns its keep. Some is bought because someone made it cheap. Know which of yours is which.
4. Keep dry powder: Evolution, not revolution. The firms holding cash going into volatility are the ones who buy the dip instead of becoming it.

The floor looks packed tonight. Just make sure you know who paid for the crowd before you build your whole set around it.

What's Coming

The Single-Supplier Era Keeps Cracking

AMD's roadmap past MI450 tells you this is not a one-off. Expect more frontier labs to sign second and third chip suppliers over the next two quarters, and expect ”which silicon” to become a board-level question, not a procurement footnote.

Custom Chips Go From Hyperscaler Toy to Enterprise Default

Databricks moving onto custom silicon is the leading edge. As GPU costs stay stubborn, more enterprise data platforms will quietly route workloads onto custom chips, and the pitch to your CFO shifts from ”best model” to ”best cost per token.”

Benchmark Trust Becomes a Purchase Term

The scrutiny over what an AI vendor actually delivers is about to harden into procurement language. Watch buyers start demanding independent evaluation evidence, not vendor-run benchmarks, before they sign.

For Your Team

Monday's meeting prompt: ”List every AI workload we run that depends on a single compute supplier. If that vendor doubled its price or ran out of capacity next quarter, what breaks, and what is our second source?”

Share-worthy stat: Broadcom just told investors its AI-chip revenue will more than double to $16 billion in a single quarter, while a rival, AMD, spent this week signing one of the biggest AI labs as a customer. The compute land grab is not slowing down.

Go deeper: Track where AI compute money is really moving, in real time →

The Track of the Day

”If the perceived AI winners become sellers or funding shorts themselves, the rotation back into software stocks could be a violent whiplash.”
From a Q2 2026 investor letter

The whole week rhymes with that line. The money is moving fast and in circles, and the operators who keep some cash and a named second supplier are the ones who get to enjoy the show instead of becoming it.

We scanned 190,000 articles this week so you don't have to. Data Pains → Business Gains.

Published: July 24, 2026 | Curated by Yves Mulkers @ Ins7ghts