Here is the uncomfortable fact that these headlines show: Even while deployments at the edge are falling apart because of trust, UX, and operational debt, AI's center of gravity is becoming more stable in a few key places.

It's hard to notice the pattern in this week's events. There is a lot of data and computational power on one side: Nvidia made $46.7 billion in sales in the second quarter, which was a 56% increase from the same time last year. They also said that only two customers made up 39% of that total. Meta's $14.3 billion deal with Scale AI is already showing signs of trouble. Meta is borrowing space from competitors to train the next generation of models. Reliance in India is putting together a national AI backbone with U.S. platforms. At the same time, AWS is rolling out industrial-grade scaffolding (Bedrock security with Datadog and HyperPod autoscaling), and Google is going deeper into AI creation tools. On the other hand, the last mile is a mess: Taco Bell's AI drive-thru doesn't work, Anthropic's opt-out deadline makes data tensions worse, and marketers are having a hard time being believed. In the fifteen years I've been monitoring this industry, I've rarely seen the stack so consolidated at the top and so weak at the point of use.

Nvidia's concentration risk is the most important story, so let's get into it. You can tell that two unnamed clients (let's call them "A" and "B") are driving roughly 40% of a record quarter. This means that a small group, perhaps hyperscalers with an insatiable need for training and inference, controls the supply chain and prices. Investors loved Jensen Huang's prediction that the global AI infrastructure market would be worth $3–$4 trillion over the next five years. However, the stock price dropped because the market is starting to wonder how long that demand can stay this concentrated and what will happen when those same buyers start buying custom silicon. Google already uses TPUs, Amazon is using Trainium, Microsoft is ramping up Maia, and Meta is making its own chips while renting space from competitors. This is an unpleasant but transitory dependency.
What people in the procurement world know: From 2024 to 2025, the economy has been based on allocating capacity, not finding prices. You grew if you could get H100/GB200 slots and HBM; if not, you waited. That gives Nvidia and TSMC a lot of power in their packaging queue and in the wallets of a few clouds. But this is exactly the situation that leads to substitution. Expect a gradual change: 1) In the short run, hyperscalers keep buying Nvidia to fulfill SLAs. 2) In the medium term, 20–30% of new training loads go to house silicon and AMD's MI3xx/MI350 class. 3) In the long term, more inference moves to optimized accelerators and NPUs at the edge. The result? Nvidia is probably still a huge source of income, but by the end of 2026, growth will slow down, the mix will change, and gross margins will meet gravity.

This focus also explains why Meta's Scale AI is shaky. Companies that sell data curation, evaluation, and agent orchestration are learning that the computing is controlled by the same platforms that are also their competitors. If Meta has to rent from a competitor to teach Llama's replacement, every step in the process (labeling, red-teaming, evaluations) becomes a negotiation between enemies. That doesn't mean tools aren't important; it just means that distribution and compute are more important than anything else. At the same time, policy is changing the data layer. Anthropic's opt-out deadline shows that default-open user data is no longer available; instead, trustworthy first-party and licensed corpora are winning. In marketing, the trust crisis is making people want guarantees of performance and proof of origin instead of demos. Meta is making it harder for teens to use chatbots, which is the same thing as putting limits on growth.
Governments are holding down the industrial base at the bottom of the stack. The U.S. arrangement that stops Intel from selling its foundry unit shows you the playbook: onshore capacity is strategic, not optional. Reliance's effort with Google and Meta to develop India's AI backbone shows that there is a new multipolar compute order: coalitions, not just one American vendor, will meet sovereign demand. If you're a new business, national clouds and compliance rules will affect your go-to-market as much as the model you choose.

Then there's the reality check on deployments. Because of accents, drive-thru acoustics, menu complexity, and antagonistic customer behavior, Taco Bell's 500-location experiment with voice ordering in the wild created viral edge cases, such as 18,000 water cups. 911 centers use AI triage for non-emergency calls, and Buenos Aires uses a Bedrock-backed assistant for government services. These are examples of restricted, high-SLA areas with specific guardrails that are working. The lesson from watching business waves for a decade and a half is that operational discipline beats ingenuity in the second year after a platform excitement peak. Datadog's Bedrock posture management and HyperPod autoscaling are boring on purpose, and that's exactly what the next stage of adoption needs.
Some people think that the data squeeze might make things better. As more people choose not to use the service and regulators are stricter, models will rely more on licensed, domain-specific, and synthetic data with better lineage. That makes things more expensive but more reliable, especially for agentic processes that will be rated on SLAs, not vibes. Companies will start putting agent SLAs into contracts in 2026.

What to do now that the market has changed: Investors should stay long AI infrastructure but hedge their concentration. Combine Nvidia exposure with AMD, Broadcom, HBM providers (SK hynix, Samsung, Micron), sophisticated packaging, liquid cooling, and power infrastructure. Add picks and shovels for security, observability, and evaluation. If hyperscalers speed up bespoke silicon production faster than consensus, expect a lot more resets.
- Startups: Don't put your life on the line by renting out limited training space. You can win by having more depth in your process, better safety tools, and integrations that incumbents can't keep up with. Make data provenance and consent part of the design. Don't make the consumer voice UX too fragile. Instead, focus on limited, high-value operations (such support triage, claims, and KYC) that have clear results.
- Businesses: From the start, talk about reserved capacity, multi-model, and multi-cloud. Give agents an identity, a policy, and an audit, just like you would with software-defined employees. Set up model registries, evaluation pipelines, and responses to incidents. Move money from testing to reliability engineering. For everyday users, there should be more guardrails, fewer strange chatbot moments, and clearer consent channels. Some AI functions will only be available to people who pay for them. Hold contractors accountable for correctness and provide them alternatives for escalation. This should make government services and healthcare discovery a lot better.

Brave predictions: Hyperscalers move 25% of new training to bespoke or alternative accelerators within a year. This slows Nvidia's data center expansion to the high teens by 2026. By the middle of 2026, "agent uptime" and "hallucination budgets" will be shown on business dashboards next to "cloud compute." And by 2027, at least three governments will say that sovereign AI stacks are important infrastructure and set restrictions for how startups can buy and train on them. The main point is that AI's power is to centralize, but it will gain credibility at the fringes. Owning trust gives you strength, but owning compute gives you time. The people who win the next cycle will do both.

AI’s Centralization Crunch: Nvidia’s Two-Customer Problem and the Agent Reckoning