TechLook
The AI race just made three things undeniable: Sora's consumer bet is over, infrastructure is the new moat, and the capital map is splitting down the middle. Here's what actually happened today — and what it means for how you build.
Sora Burned $1M a Day. Then It Died.
OpenAI's consumer video generator shuts down today — nine months after launch — having never escaped the compute trap that was always waiting for it. The app peaked at roughly one million users, dropped to under 500,000, and revenue never came close to covering its burn rate. Disney's character licensing deal ends with the app. API developers have until September 24 to migrate.
Here's everything you need to know:
- Sora launched in December 2025 with waitlists in the hundreds of thousands
- The consumer app burned an estimated $1M per day in compute — confirmed by multiple reports
- Disney character licensing was a significant revenue stream; its removal suggests the deal wasn't renewable
- Runway Gen-4.5, Kling 3.0, Google Veo 3.1, and ByteDance's Seedance 2.0 are positioned to pick up displaced creators
- The API survives — enterprise customers who integrated Sora programmatically keep access
- OpenAI is shifting resources toward coding, agentic AI, and enterprise products — the pivot is now concrete
- Competitors have caught up: practitioners rate Runway and Kling as on par or better for many use cases
Sora's collapse is the clearest data point yet that consumer AI generation can't make the unit economics work yet. The enterprise pivot isn't a strategy slide — it's survival. For builders, this is a useful rule: demos get attention, but infrastructure at cost determines which products survive.
The AI Race Just Became an Infrastructure Race
This week's Google Cloud Next generated over 260 announcements — but they reduce to one sentence: the industry moved from "can we build AI agents?" to "how do we manage thousands of them?" That second question is harder, more expensive, and more productively interesting.
Here's everything you need to know:
- Gemini Enterprise Agent Platform: Build, scale, and optimize fleets of agents with a no-code Agent Designer and support for 200+ models
- Workspace Intelligence: Connects Gmail, Docs, Sheets, Slides, and Drive into a unified knowledge graph — every tool you already use, wired together
- Apple + Gemini for Siri: Gemini becomes the cognitive engine behind Apple's assistant overhaul, with a standalone app, persistent chat logs, and a full chatbot experience
- New TPUs: TPU 8t for training, TPU 8i for inference — the hardware layer everything runs on
- The management layer is the product now. This is where enterprise money flows.
If you're building agentic products, Google is signaling where enterprise budgets will go: not the model, but the orchestration layer. Who manages the agents? Who monitors them? Who fixes them when they behave unexpectedly? These are the product questions that matter to procurement teams today.
DeepSeek V4 Benchmarks Near GPT-5.4 — at One-Sixth the Price
DeepSeek V4 is out, and it benchmarks near GPT-5.4 and Gemini 3.1-Pro on reasoning tasks — at roughly one-sixth the cost. Pricing sits at $1.74 per million tokens versus GPT-5.5's $5 and Opus 4.7's $5 for comparable performance tiers. This isn't a toy model. It's a legitimate competitor at a structural price advantage.
Here's everything you need to know:
- V4 Pro and V4 Flash both ship with 1M-token context windows
- Huawei Ascend chip support is confirmed — a viable domestic Nvidia alternative that works despite export restrictions
- US State Department formally accused DeepSeek and other Chinese labs of IP theft from American companies — this launch lands in the middle of escalating US-China AI rivalry
- V4 Pro benchmarks near GPT-5.4 and Gemini 3.1-Pro on reasoning tasks
- The argument for Chinese AI infrastructure independence is no longer theoretical. Huawei Ascend is production-ready.
The cost argument is real but not the whole story. The Huawei angle is the bigger signal: Chinese AI infrastructure is becoming independent of Western chips. That's a supply chain shift with long-term pricing implications globally. If you're building for global markets, the AI stack looks different in China than it does in the US — and that gap is closing.
Your AI Agent Might Be Getting You Worse Deals Than You Think
Anthropic ran an internal experiment called "Project Deal." Sixty-nine employees each got a $100 gift-card budget to buy items from coworkers. AI agents did the negotiating. The results reveal something uncomfortable about what happens when AI represents you in markets you can't fully audit.
Here's everything you need to know:
- 186 deals completed, worth over $4,000 across four marketplace setups
- Users with Opus agents fetched $3.64 more on average for identical items — same product, same instructions, better outcome
- Users with Haiku agents bought at higher prices and sold at lower ones, consistently
- Fairness ratings were nearly identical between both groups: 4.06 vs 4.05 — neither side noticed the gap
- 46% of participants said they would pay for the agentic negotiation service
- Policy and legal frameworks for agent commerce don't exist yet
The sharp finding: once agents negotiate on your behalf, unfairness becomes harder to detect because the work happens offscreen. A marketplace can appear calm while weaker buyers quietly pay more. Better model, better outcome — and the user sees no signal that this is happening. For founders building agentic commerce products: the fairness gap is your liability. If your product systematically advantages users with access to better models, you will face regulation, backlash, or both.
Google Commits $40B to Anthropic. Here's What It's Actually Buying.
Google is committing up to $40B to Anthropic: $10B now, another $30B tied to performance targets, at a reported $350B valuation. Alongside the investment, Google is providing 5GW of cloud compute — the actual industrial capacity to run frontier models at scale.
Here's everything you need to know:
- $10B deployed immediately; $30B contingent on Anthropic hitting specific performance targets
- Anthropic's Mythos model has major cybersecurity potential but is access-limited due to misuse risks
- The AI race is shifting from model quality to compute infrastructure — cloud capacity, TPUs, and energy supply are the strategic weapon now
- Google isn't just investing money — it's providing the cloud, chips, and TPU capacity Anthropic needs to compete
- The $40B reflects what it costs to run frontier AI at scale, not just to develop it
The AI race is becoming less about clever demos and more about who controls the heavy industrial infrastructure behind them. For builders, this means the real moats in AI aren't model quality — they're access to compute, energy, and infrastructure. That's where the leverage sits.
Apple Just Put Google's AI in Siri — And Both Companies Needed It
Apple and Google are doing something unusual: a deep partnership that positions Gemini as the cognitive engine behind Siri's overhaul. The new Siri will have a standalone app, persistent chat logs, and a full chatbot experience — a completely different product from the voice assistant founders have known for a decade.
Here's everything you need to know:
- Gemini powers the new Siri experience across iOS
- Standalone app, persistent memory across sessions, full chatbot interaction — a fundamentally different product
- Apple stops being a voice command interface and becomes a conversation platform
- Apple picks up Gemini's reasoning capabilities; Google picks up Siri's distribution
- Both companies are betting their AI future on a partnership that would have been unthinkable two years ago
For founders building AI products: Apple embedding Gemini means Gemini becomes the default AI experience for hundreds of millions of iPhone users. Distribution through platform partnerships is becoming as important as model quality. If you're building AI-native products, the platform relationship matters as much as the model.
China Is Blocking US Money From Its AI Startups
Beijing is cutting off what it sees as quiet exit routes for sensitive AI technology. The government is blocking top companies — including Moonshot AI and StepFun — from taking US money without explicit government approval. ByteDance secondary share sales may face the same restrictions. The trigger: Meta's reported $2B investment in Manus.
Here's everything you need to know:
- Moonshot AI and StepFun cannot raise from US investors without explicit government approval
- ByteDance secondary share sales may face the same restrictions
- Meta's $2B investment in Manus is cited as a trigger — capital, ownership, and frontier AI talent被视为出口渠道
- Chips, models, data, and ownership are now part of the same security fight
- Chinese AI startups can still raise huge rounds, but the buyer list is shrinking
- The AI capital map is splitting — and the split is accelerating
For founders with any cross-border AI ambitions: the regulatory environment around AI investment is hardening on both sides. This isn't just about chips anymore. Capital flows, ownership structures, and talent movement are all being reviewed. Build accordingly.
Cohere Acquires Aleph Alpha for ~$20B — Sovereign AI Is Now a $20B Market
Germany's AI lab Aleph Alpha has been acquired by Cohere in a deal reportedly valued around $20B. The target is clear: governments and enterprises in Europe that want a sovereign AI option outside the control of American or Chinese platforms.
Here's everything you need to know:
- Aleph Alpha built its reputation on European data sovereignty and government partnerships
- Cohere's enterprise focus meets Aleph Alpha's government credibility
- $20B valuation for a European AI lab signals real institutional demand for non-American AI infrastructure
- For founders building in regulated industries — finance, healthcare, government — this deal is a data point on who the serious buyers are
Sovereign AI isn't a niche concern anymore. It's a $20B market signal.
The Cursor Is Becoming an Operating System Layer
Clicky, the AI-powered cursor, can now spawn sub-agents to complete tasks autonomously. It controls native Mac apps and builds custom software tools on the fly — moving from a cursor to a programmable construction layer for software.
Here's everything you need to know:
- Sub-agent spawning means a single prompt can trigger a cascade of autonomous work
- Native Mac app control and custom tool building on the fly from a single prompt
- Use cases expanding: research, file organization, custom tool construction
For developers: the abstraction layer is shifting. Your tools are only as useful as the agentic layer built on top of them.
⚡ Quick Hits
- xAI + Cursor + Mistral: In talks about a three-way partnership, with SpaceX's $60B option-style structure signaling the deal's ambition.
- OpenAI: Clarified law enforcement protocols after earlier criticism over its shooting incident response.
- Maine: Governor vetoed a statewide data center moratorium, exempting a locally supported project in Jay.
Techlook — AI & tech signal for founders and builders.