// AI Tangle
The Week AI Got a Balance Sheet
Nvidia's Wednesday print, Stripe's $7 billion routing bet, DeepSeek's price hike, and OpenAI's IPO year — the money side of AI grew up this week.

I've been watching tech earnings since the days when Cisco was the most valuable company on Earth because it sold the plumbing for the internet, and this week has that same feel. Nvidia reports Wednesday against a $91.85 billion consensus, while the four biggest hyperscalers spent $166 billion on capex in a single quarter to make numbers like that possible; meanwhile, Stripe is acquiring OpenRouter for more than $7 billion, Google has handed its Agent2Agent protocol to the Agentic AI Foundation alongside MCP, DeepSeek's V4 lineup comes with prices up as much as 1,100%, and OpenAI CFO Sarah Friar says the company could be public by 2027 or sooner. The AI story this week isn't about another model release—it's about the balance sheet. Join us at AI Tangle as we untangle this week's biggest happenings in AI!
// The Big AI Story
Wednesday's Nvidia Print Is a Referendum on $166 Billion a Quarter
Nvidia reports second-quarter fiscal 2027 results Wednesday, August 26, after the close. Consensus sits at $91.85 billion in revenue, up 96.5% year over year, with earnings expected around $2.09 per share — roughly double a year ago. Data center revenue was 92% of the company last quarter. At this point, Nvidia isn't a chip company with a data center business. It's a data center business with a gaming hobby.
The number behind the number is hyperscaler capex. Amazon, Alphabet, Microsoft, and Meta spent a combined $166 billion on capital expenditures in Q2 — up 27% sequentially and 87% year over year. That's the demand side of Nvidia's income statement, and it's the single line every CFO watching this print actually cares about. When four companies are spending at a two-thirds-of-a-trillion-a-year pace on infrastructure, the money has to come back from somewhere — and "somewhere" is the software bills the rest of us pay.
The real tell Wednesday won't be the quarter — it'll be the guide. Consensus expects Nvidia to point to roughly $103 billion for Q3, which implies accelerating sequential growth at a scale no company has ever sustained. If the guide lands, the capex machine keeps running through 2027. If it doesn't, every AI budget conversation in corporate America changes tone by Thursday morning.
Here's the business read: you don't own Nvidia's problem, but you rent it. Vendor pricing, token costs, and every "AI included" line item in your renewals are downstream of whether this spending earns a return. Wednesday at 5pm ET is the closest thing this industry has to a quarterly weather report. Watch the guide, not the beat.
// 4 Quick Hits
1. Stripe Is Buying OpenRouter for More Than $7 Billion
Stripe is acquiring OpenRouter, the gateway that routes API calls across hundreds of AI models, in a deal Bloomberg reports at over $7 billion — more than 5x OpenRouter's valuation from three months ago. The logic: every model call is a metered transaction, and Stripe already runs the metering business for the internet. For buyers, the read is simple — model routing is consolidating into financial infrastructure, and the neutral middle layer you use to avoid vendor lock-in now has an owner.
2. Google's A2A Protocol Gets a Standards Body
Google moved its Agent2Agent protocol — the standard that lets AI agents from different vendors talk to each other — to the Agentic AI Foundation, the group that already stewards MCP. The foundation has grown from under 40 members at its December launch to more than 250, including Google, Microsoft, Amazon, Anthropic, OpenAI, and Bloomberg. As the AAIF's executive director put it: "Companies don't want just one protocol; they want the whole stack to be open." For enterprises, this is the boring news that matters — agent interoperability now has one address instead of three.
3. DeepSeek's V4 Lineup Comes With a 1,100% Price Increase
DeepSeek shipped its V4 model lineup and rewired its API pricing — rates up as much as 1,100% from the old list, with new off-peak pricing at exactly half the peak rate, effective August 16. The company that spent 2025 as the industry's price floor is now billing like a utility, complete with time-of-day rates. If your AI budget assumed cheap Chinese inference forever, this is your notice — and if you have batch workloads, it's also your first real arbitrage opportunity: schedule them off-peak and cut that line item in half.
4. OpenAI's CFO Names the IPO Year: 2027
OpenAI CFO Sarah Friar told employees the company "will be a public company in 2027" or sooner, pushing past earlier hopes for a fall 2026 listing. The backdrop: a revenue run rate that just topped $40 billion, roughly double a year ago, a confidential S-1 already filed, and an $852 billion private valuation to grow into. For enterprise buyers, a public OpenAI means something you've never had from your AI vendor: audited financials and quarterly disclosure. You'll finally get to read their 10-K the way I keep telling you to read everyone else's.
// 3 AI Tools
Notion — Notion's August updates are quietly agent-forward: a model picker with scorecards comparing speed, intelligence, and cost per model, agents that trigger automatically when AI meeting notes finish, and a new developer portal for managing agent connections and tokens. If your team already lives in Notion, you now have an agent platform you didn't have to procure.
OpenAI Presence — OpenAI's enterprise platform for production voice and chat agents, with policies, guardrails, and escalation rules built in. OpenAI runs its own phone support on it and says it resolves 75% of inbound issues without a human. Limited availability, but this is the shape of enterprise support in 2027.
Cloudways Managed AI Agents — DigitalOcean's Cloudways now offers managed hosting for open source agents like OpenClaw and Hermes. If you're a smaller shop that wants agents without an infrastructure team, this is the accessible on-ramp — the managed-WordPress model, applied to agents.
// The Extra Read
Bessemer's atlas piece on agent security is the sober companion to everything above. The week's news is all about agents getting cheaper, more standardized, and easier to deploy — this is the essay on what happens when they're deployed faster than they're secured. Read it before your next architecture review, not after your first incident.


