The Great Infrastructure Squeeze: When Cloud Capacity Runs Out and Models Commoditize

Today's AI landscape is revealing a brutal paradox: just as frontier models are becoming commoditized and affordable, the infrastructure to run them is completely sold out. AWS is booked through 2028. Chinese labs are shipping open-weight models that match GPT and Claude. And the entire digital publishing industry is watching Google's AI summaries delete their business models overnight.

Welcome to the infrastructure squeeze — where the picks and shovels cost more than the gold, and yesterday's moats are today's swimming pools.

1. AWS Sold Out Through 2028, Jassy Confirms

Amazon Web Services capacity is essentially booked through 2027, with 2026 already completely sold out, CEO Andy Jassy revealed on an a16z podcast. AWS is only holding back minimal capacity for on-demand and inference workloads. This represents the most extreme capacity constraint in cloud computing history.

Let's be clear about what "sold out" means: if you're a startup that needs significant GPU clusters for training or fine-tuning, and you didn't lock in a multi-year contract 12-18 months ago, you are functionally locked out of AWS until 2028. Not "it'll cost more" — you literally cannot buy what you need.

This isn't AWS being conservative with capacity allocation. This is a fundamental mismatch between AI demand growth (exponential) and datacenter construction timelines (3-4 years from planning to production). The hyperscalers saw the AI wave coming in early 2024, started construction, and still can't build fast enough.

🔥 Spark's Hot Take: The real story isn't that AWS is sold out — it's that cloud pricing power is about to swing violently back to the hyperscalers after a decade of competition keeping costs down. Remember when Cloudflare and DigitalOcean forced AWS to cut prices? That era is over. If you're building AI infrastructure into your product and don't have locked-in pricing through 2027, your unit economics are about to get destroyed. The only escape hatch is smaller providers like CoreWeave or Lambda Labs, and even they're feeling the squeeze as their costs rise.

The winners here are obvious: enterprises that signed reserved instance contracts in 2024-2025 at pre-squeeze prices. The losers are late-stage startups trying to deploy models in production right now, suddenly unable to get GPU clusters at any price. Your AI roadmap just got a lot more expensive — or impossible.

2. Moonshot's Kimi K3.1 Pro Matches Opus Performance

Chinese AI startup Moonshot AI launched Kimi K3.1 Pro, scoring 55 on Artificial Analysis — tied with frontier models and ahead of GPT-6 Astra and Gemini 4 Argon (both at 53). The model costs $2/million input tokens and $10/million output, half the price of Claude Opus 5.5. Full model weights will be released by November 4th, making this the first open-weight model to reach frontier-level performance.

Read that last sentence again: the first open-weight model to reach frontier-level performance. That's the sound of the AI model moat cracking in half.

For years, the implicit assumption was that only companies spending $10 billion+ on training runs could produce frontier models. OpenAI, Anthropic, Google, and DeepSeek were running a capital-intensive race that smaller players couldn't compete in. Then Moonshot shows up with a model that matches Claude Opus 5.5 on benchmarks, costs half as much, and will be fully open-weights in two weeks.

If you're paying $4/million tokens for Claude Opus API calls right now, you have a credible alternative at $2/million that you can also run on your own infrastructure. That's not a minor improvement — that's a different economic model entirely.

🔥 Spark's Hot Take: This is the moment China's AI ecosystem stopped playing catch-up and started competing on equal footing. Not with closed APIs and data center lock-in, but with open weights you can download and deploy yourself. Every AI-powered SaaS company that built their entire strategy around "OpenAI has the best models" just got a reality check about commoditization risk. The frontier is no longer a proprietary advantage — it's a temporary head start that lasts 6-12 months before open alternatives catch up. If your competitive moat is "we use GPT-6," you don't have a moat.

The strategic question for founders: do you keep paying premium prices for slightly-better closed models, or do you invest in the infrastructure to run open-weight models yourself? Because in 18 months, that choice might not exist — everyone will be running open models, and the differentiation will be in your data, fine-tuning, and application layer.

3. OpenAI Price Cuts Flip Rippling From Anthropic

Rippling engineers shifted from 57% Anthropic/16% OpenAI in May to 18% Anthropic/52% OpenAI today, driven entirely by OpenAI's aggressive price cuts. CEO Parker Conrad confirmed the reversal was about cost reduction, not model quality. The swing happened over just five months as OpenAI systematically undercut Anthropic's pricing.

This is the first concrete evidence that OpenAI's pricing war is working. Anthropic spent years cultivating a brand around "constitutional AI" and safety, got Rippling's engineering team to standardize on Claude, and then lost them in five months because OpenAI went cheaper.

Think about what that means: Rippling didn't switch because Claude got worse or GPT got dramatically better. They switched because OpenAI cut prices and engineering teams are rational economic actors. Model quality matters, but when the gap is narrow enough, cost wins.

Anthropics's nightmare scenario is playing out in real-time. You can't compete on price against a company backed by Microsoft's balance sheet. You can't compete on quality when the models are functionally equivalent for most tasks. And you can't compete on switching costs when changing from Claude to GPT requires changing about eight lines of code.

What's Anthropic's move here? They can't match OpenAI on price without burning through their runway. They can't meaningfully differentiate on safety when customers don't pay extra for it. Their best bet is probably to focus on specific verticals (healthcare, legal, finance) where compliance and safety actually matter enough to justify premium pricing — but that's a much smaller TAM than "general-purpose AI for all developers."

For everyone else, the lesson is stark: API-based AI infrastructure has zero switching costs. If your competitive advantage is "we use the best model," you have no competitive advantage. The moat has to be in your data, your fine-tuning, your UI/UX, or your distribution — not in which API you call.

4. Instacart Exploring Sale to DoorDash or Uber

Instacart is in active sale talks after 14 years as an independent company, currently valued at $12 billion. The grocery delivery app has held at least preliminary conversations with DoorDash and Uber, though talks with Uber have stalled. Shares jumped 10% on the news, closing at $56.44.

Let's trace the arc: Instacart raised huge rounds during the pandemic, hit a $39 billion valuation, went public in September 2025 at a reduced but still hefty valuation, and is now shopping itself at $12 billion — a 69% haircut from peak.

That's not a unique story. That's the entire consumer startup playbook from 2020-2026: raise at nosebleed valuations, IPO into hype, watch public market multiples compress as growth slows, get acquired by a larger platform player.

For DoorDash or Uber, buying Instacart is about vertical integration and market consolidation. DoorDash especially makes sense — they're already doing restaurant delivery, adding groceries creates a fuller on-demand delivery platform, and they eliminate a competitor in one move. For Uber, it's less clear-cut since they're increasingly focused on autonomous vehicles and robotaxis, but owning the full delivery stack still has strategic value.

The deeper question is what this says about consumer tech exits. Instacart should have been a generational success story — a pandemic winner that built real infrastructure and went public. Instead, it's getting sold for parts three years later. If you can't defend margins against Amazon's logistics network and Walmart's scale, your path to independence gets narrower every quarter.

The real winner might be DoorDash, which gets to consolidate on-demand delivery at a reasonable price while Uber chases robotaxis and Amazon focuses on Prime. Sometimes the best strategy is just to survive longer than your competitors and buy them when their options run out.

5. Google's AI Overviews Devastate Publisher Traffic Overnight

Google's AI-powered search summaries have fundamentally broken the reciprocal relationship between Google and online publishers. Sites that spent years optimizing for SEO are watching traffic collapse as AI Overviews answer queries directly, eliminating the need to click through to source articles. Publishers are seeing traffic declines of 40-60% in some categories.

This is the beginning of the Great Unbundling of the web, and it's going to be ugly.

For 20 years, Google and publishers had a deal: publishers created content and optimized it for Google's algorithm, Google sent them traffic, everyone made money from ads. That ecosystem generated hundreds of billions in value and employed hundreds of thousands of people.

Then Google launched AI Overviews and flipped the table. Now when you search for "how to fix a leaky faucet," you don't click through to Home Depot's blog post — you read Google's AI-generated summary that extracted the information from that post without sending any traffic. The publisher gets zero clicks, zero ad impressions, zero revenue.

Google's response, essentially, is "we're still sending some traffic, and AI Overviews help with complex queries." That's PR spin. The traffic drops are real, they're severe, and they're permanent. If Google can answer the query with an AI summary, it will — because that keeps users on Google.com where Google controls the ads.

What's a publisher supposed to do? The smart play is what it's always been when platform risk materializes: build direct relationships. Email lists. Paywalls. Apps. Communities. Anything that doesn't depend on Google deciding to send you traffic. The entire SEO industrial complex — the consultants, the tools, the "growth hackers" — is about to become a legacy industry, like Yellow Pages salespeople.

The deeper implication is that the open web as we knew it is dying. Not because people stop publishing, but because the economic model that funded most publishing (ad-supported content optimized for search traffic) just got obliterated. What replaces it is unclear, but it probably looks more like Substack and Patreon than BuzzFeed and HuffPost.

Bottom Line

The infrastructure squeeze is here: cloud capacity is sold out through 2028 while frontier models are commoditizing to half-price open-weights, incumbents are slashing prices to defend market share, and platform risk is materializing across the stack from cloud providers to search engines. The winners will be companies that locked in infrastructure early, built real moats beyond "we use the best API," and own direct customer relationships instead of renting them from Google. The question every founder should ask this week: if your primary cloud provider, model provider, and traffic source all disappeared tomorrow, would you still have a business?