AI Week in Review: New Models, Billions, and an Ethics Exit

This week captures the industry’s split personality rather well: faster, cheaper models on one side, a financing scheme for AI infrastructure that borrows straight from commercial real estate on the other, plus an uncomfortable security flaw in chatbots’ reasoning processes and a quiet leadership change over who at OpenAI still minds ethics. Five stories that, taken together, say more about the state of the industry than any one of them alone.

Key takeaways

• Google released Gemini 3.7 Flash just three weeks after its predecessor, beating Claude Sonnet 5 and GPT-5.6 Terra on code quality by its own benchmarks, at half the price.
• Nvidia lined up more than $500 billion in third-party capital with six Wall Street partners to finance AI data centers as their own asset class.
• A research team showed that the supposedly encrypted internal reasoning of GPT-5.6, Claude, and Gemini can partly be read in plain text.
• OpenAI’s only dedicated ethics lead, Chloé Bakalar, left the company after eleven months; the role remains unfilled.
• The two-month-old startup River AI raised $1.1 billion despite offering little more than a narrow cloud API so far.

Gemini 3.7 Flash: Google turns release speed into strategy

Google shipped Gemini 3.7 Flash just three weeks after version 3.6. According to Google, its FrontierCode score jumped from 34.4 to 43.6 percent, and DeepSWE from 49.0 to 65.3 percent, gains the company attributes to algorithmic improvements rather than added compute. DeepSeek made its V4-Pro model generally available at the same time, with a peak/off-peak pricing structure reminiscent of electricity tariffs. What matters here is less the individual model than the pace: when a flagship update lands every three weeks, the model name itself loses meaning, what counts is how a vendor stacks up in any given month. For users, the takeaway is to keep re-checking price-to-performance rather than committing to one model, as our analysis of Gemini’s user growth already showed, benchmark percentage points rarely decide the matter alone.

Nvidia turns data centers into an asset class

With Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs, and KKR, Nvidia has lined up six of the world’s largest asset managers for a joint financing platform meant to mobilize over $500 billion in third-party capital for AI data centers, power hookups, and hardware. The trick: the money never touches Nvidia’s own balance sheet, instead treating compute capacity like commercial real estate or toll roads, an asset investors can leverage and refinance. CEO Jensen Huang accordingly called chips an “investable asset” on CNBC. Exactly how the money flows, and who’s left holding the bag if the multi-billion-dollar bets on future AI demand don’t pan out, is something we’ve broken down in detail.

Security flaw: AI reasoning traces weren’t so encrypted after all

A research team led by Alexander Panfilov disclosed a flaw affecting all three major vendors: GPT-5.6, Claude Opus 4.8 and Haiku 4.5, and Gemini all encrypt their internal reasoning, but use the same global key per model family instead of one bound to a session or user. From 6,700 public agent logs, the researchers extracted 62 API keys and 33 passwords, some of which appeared only in the internal reasoning, never in the visible answer. Decoding 10,000 traces cost roughly $720 by the team’s own estimate, hardly a price that deters a serious attacker. All three vendors confirmed the findings after responsible disclosure and patched the issue. For companies running agents in production, the lesson is simple: encryption alone is no security guarantee when the key is the same for everyone.

OpenAI loses its only ethicist, quietly

Chloé Bakalar joined OpenAI from Meta in August 2025 as the company’s sole dedicated lead for ethical questions in model development. She left again in July 2026, with no public announcement and no successor; her responsibilities now spread across several research teams. Her departure joins a pattern: Safety Systems lead Johannes Heidecke, Chief Futurist Joshua Achiam, and earlier robotics lead Caitlin Kalinowski all left OpenAI in 2026. A company spokesperson told media that ethics “doesn’t live with one owner or team,” which may be true organizationally but says little about who now has veto power when a product decision raises ethical concerns. We’ve covered the personnel move and its context in a separate piece.

$1.1 billion for a two-month-old startup

Igor Babuschkin, xAI co-founder and formerly a researcher at DeepMind and OpenAI, founded River AI in June 2026 and had already raised $1.1 billion in a combined seed and Series A round by August, led by General Catalyst, with Nvidia, AMD Ventures, Y Combinator, and Temasek also participating. The business behind it: a cloud API that lets companies fine-tune open models like Qwen3.6, Kimi K2.6, or GLM 5.2 through reinforcement learning and LoRA, paired with a vision in which AI agents are “owned” by people and trained for them, rather than merely rented. A valuation of up to $5 billion for a company with no meaningful track record shows one thing above all: in the current funding climate, the founding team’s name often counts for more than the product. More on the company and its business model is in our full analysis.

Bottom line: speed, capital, and open questions

What connects these five stories is a pattern: models keep getting faster and cheaper, the capital behind them is growing to a scale that borrows straight from real estate finance, and basic questions of safety and accountability remain unresolved, whether that’s reasoning traces that turned out not to be encrypted after all, or an ethics post nobody bothers to refill. If there’s one lesson from this week, it’s that speed is no substitute for care, not in model training and not in corporate governance.