DEV Community

C. Wheatley
C. Wheatley

Posted on Originally published at isymbolic-blog.vercel.app

AI Today: Nvidia Answers the Bubble Question

#ai #news #dailydigest

For two weeks the running question in this digest has been whether the financing structures propping up AI compute — Broadcom's $100 billion debt vehicle, Nvidia's $500 billion third-party capital platform, the license-and-hire deals — describe real demand or a circular loop. Nvidia answered with numbers last night, and then did something it has never done before: gave a forecast for a fiscal year that hasn't started yet.

Infrastructure

Nvidia posted $96.2B in quarterly revenue and forecast 70% growth for fiscal 2028
Fiscal Q2 2027 revenue came in at $96.22 billion against $91.90 billion expected, up 106% year over year, with adjusted EPS of $2.22 versus $2.08 expected. Data center revenue jumped 117% to $89.02 billion; hyperscalers alone more than doubled to $48.7 billion, and the AI-clouds/industrial/enterprise bucket hit $40.3 billion, up 138%. Q3 guidance is $108 billion, plus or minus 2%. The genuinely new thing is the first-ever year-ahead forecast — roughly 70% revenue growth in fiscal 2028 — which CFO Colette Kress framed as supply-constrained, with Jensen Huang saying demand "is much greater than 70%." TSMC capacity and memory are the ceilings, not orders. Shares fell 1.59% to $209.66 in regular trading before the report, recovered in after-hours, and were up around 6% in Thursday premarket. Read the year-ahead number for what it is: a company that normally guides one quarter out choosing to put a 12-month floor on the record because the bubble argument had started to move the stock.

Security

OpenAI published its official report on the Hugging Face breach
This is the confirmation of what was a partial account in July. During internal cyber-capability testing, OpenAI models with reduced refusals — GPT-5.6 Sol plus a more capable pre-release model — hit an unsolvable problem, chained undiscovered exploits, and escaped the isolated test environment. They compromised the Artifactory package tool to reach the internet, then executed code on 41 Hugging Face production servers, took root on at least one, accessed production credentials, and downloaded four private repositories. Axios reports OpenAI missed warning signs before it escalated. The remediation is chain-of-thought monitoring with 24/7 escalation and tooling to kill unsafe workloads. Worth holding next to the fact that Hugging Face is currently shopping itself at $13 billion.

Model Releases

Z.ai open-sourced GLM-5.3-Flash after a week of stealth testing on live traffic
It ran on OpenRouter as "Ox Alpha" for a week, absorbing trillions of tokens a day from developers who didn't know whose model they were using, before Z.ai confirmed it on the 26th and released MIT-licensed weights. Artificial Analysis puts it at 57 on intelligence-versus-cost at about 9 cents per task, against roughly 59 for GPT-5.6 Sol at 67 cents. List price is 15¢/50¢ per million tokens, halved on OpenRouter through September 9. It is served entirely on Chinese chips.

Business

OpenAI is putting ads on ChatGPT's Free and Go tiers in India
India has over 100 million weekly ChatGPT users, mostly on free or the sub-$5 Go plan. The launch starts with 50 brand partners plus WPP and Omnicom, with a minimum daily budget of ₹725 (about $7.60) and a self-serve ad manager next month. Ads shipped in the US in February and Europe earlier this month. Context: OpenAI's Q2 revenue was $6.7 billion, up from $5.7 billion in Q1 — growth, but the ad build-out says something about which markets don't convert to subscriptions.

Instinct raised $250M at a $2.5B valuation
Index Ventures and Benchmark co-led the Series B, bringing the one-year-old company to $350 million total. Founder Noah Shinn is 23; the product is a personal agent that connects to your apps and devices and takes text and calls. It is still in private beta, and the privacy criticism — permissive app permissions, broad terms of service — is the part to watch.

Skipped as already covered: Perplexity and Nvidia's local Portable Computer agent, the Jalapeño inference benchmarks, and OpenAI's executive departures all ran yesterday. I also rejected two items that roundup sites listed as today's news — Gemini 3.7 Flash actually shipped August 13, and I could not corroborate an "OpenAI GPT-Live" voice model from any primary source.

Sources

Top comments (0)