Daily AI News Roundup – July 24, 2026
OpenAI is building a city-scale data center in Georgia. Google beat earnings and still got punished. A critical developer deadline hits at midnight. And the EU’s most consequential AI law kicks in next week. Here’s everything you need going into Thursday.
Want more AI news without waiting for me to finish yelling at the latest development? Check out The SkyNet Tribune, my continuously updated AI news page featuring fresh headlines from across the web every hour. It is faster than a daily roundup, less exhausting than social media, and only slightly more likely to become self-aware.
1. OpenAI Is Building a $30 Billion Data Center in Georgia — Its First Self-Built Campus
OpenAI announced Project Camellia on Wednesday: a $30+ billion AI data center campus in Effingham County, Georgia, near Savannah — the first facility the company will design and build entirely on its own. The campus has already secured a contract with Georgia Power for 3.2 gigawatts of electricity at full build-out, enough to power roughly 2.4 million US homes. Initial capacity comes online in 2028, with full development extending through 2032.
The scale puts Project Camellia in a category of its own among US AI infrastructure projects. OpenAI has simultaneously raised its projected compute capital expenditure through 2030 from $600 billion to $750 billion — a 25% increase. The company is also committing to fund the required grid infrastructure itself, a deliberate move to defuse the politically potent objection that ordinary ratepayers foot the power bill for some of the most profitable companies in history.
2. OpenAI Launches “Presence” — An Enterprise Agent Platform with Built-In Guardrails
Alongside the data center news, OpenAI officially launched Presence, a new enterprise platform for deploying AI agents across customer support, sales, and internal operations via voice and chat. Unlike raw API access, Presence wraps OpenAI’s models in company-defined policies, guardrails, and escalation rules — keeping agent behavior predictable even as products and customer patterns change.
The performance claims are striking: OpenAI says Presence met or exceeded human support-staff quality benchmarks within weeks of deployment and now resolves roughly 75% of inbound customer issues without human intervention. The catch: it’s not self-serve. Rollout is limited to enterprise customers through OpenAI’s Forward Deployed Engineers — the same high-touch model the Northslope acquisition was designed to scale. The Register notes this is OpenAI trying the consulting path — and charging consulting-firm prices for it.
3. Developer Alert: DeepSeek’s Legacy API Aliases Die Today at 15:59 UTC
If you or your team are running any automated pipeline, agent workflow, or background process that calls deepseek-chat or deepseek-reasoner — those model names stop working today at 15:59 UTC. DeepSeek is retiring both legacy aliases as part of the V4 stable release, and any call using the old names will return an error after the cutoff.
The migration itself is straightforward — same base URL, same API key, same request structure. Just swap deepseek-chat → deepseek-v4-flash and deepseek-reasoner → deepseek-v4-pro. One note: thinking mode has moved from a model name to a request parameter, so any code that relied on the old reasoner quirks needs a small cleanup. If you haven’t migrated yet, now is the time.
4. Google Beat Q2 Earnings Handily — Then Its Stock Dropped 7%
Alphabet’s Q2 report was, by almost any traditional measure, outstanding. Revenue of $119.8 billion beat expectations by nearly $3 billion, up 24% year-on-year. Google Cloud surged 82% to $24.8 billion, with a backlog now totaling $514 billion. Search grew 17%. Operating margin held near 36% for a 12th consecutive quarter of double-digit growth.
And yet: GOOGL shares fell 7% after hours as Alphabet raised its 2026 capex forecast to a midpoint of $200 billion — up from $185 billion just last quarter — and warned that 2027 spending would be “significantly” higher still. The market isn’t questioning whether AI is working for Google; it’s questioning whether the tab will ever stop growing. For Alphabet, the answer seems to be: not anytime soon.
5. The EU AI Act’s Biggest Deadline Is 9 Days Away — and 78% of Companies Aren’t Ready
Mark your calendar: August 2, 2026 is when the EU AI Act’s obligations for high-risk AI systems take full effect. That means conformity assessments completed, technical documentation finalized, CE marking affixed, and EU database registration done — for any AI system used in employment, healthcare, financial services, education, critical infrastructure, law enforcement, or judicial processes. The law’s extra-territorial reach mirrors GDPR: if your system affects EU residents, you’re in scope regardless of where you’re headquartered.
The compliance picture is alarming. As of April 2026, 78% of organizations had not taken meaningful steps toward compliance — and the fines are steeper than GDPR, reaching 7% of global annual turnover (capped at €35 million). With nine days left, the next week is going to be a very expensive one for a lot of legal departments.
6. Kimi K3’s Full Open Weights Drop Sunday — and the White House Is Already Watching
China’s Moonshot AI is set to release the full open weights of Kimi K3 on July 27 under a Modified MIT license. K3 packs 2.8 trillion parameters and a 1-million-token context window, making it the largest open-weight model ever released — and with open weights, anyone can host, fine-tune, or modify it. Independent benchmarks from Artificial Analysis rank it fourth among 189 tested models, behind Fable 5 and two GPT-5.6 Sol configurations, and ahead of Claude Opus 4.8.
The release lands in a charged political environment. The White House this week accused Moonshot of using Anthropic’s Fable model to build K3 through covert industrial-scale distillation, with Treasury threatening sanctions. Moonshot has not responded publicly. Whether open-sourcing the weights helps or hurts Moonshot’s case with Washington — by enabling transparency into the model’s architecture — is an open question. Either way, Sunday’s release will instantly put frontier-level open-weight AI in the hands of any developer on Earth.
7. Google Confirms Gemini 4 Pretraining Has Begun
Buried in this week’s news but worth flagging: Google has confirmed it has started pretraining for Gemini 4 — the next-generation model beyond the 3.x series. The announcement came the same week Google shipped three new Gemini Flash variants (3.6 Flash, 3.5 Flash-Lite, and the government-restricted 3.5 Flash Cyber) while continuing to delay Gemini 3.5 Pro with no new date given.
Pretraining at frontier scale takes months, so Gemini 4 is not imminent — but the confirmation signals Google is already building past its current generation even as Pro remains unreleased. The broader pattern: Google is shipping aggressively on the Flash (fast, cheap) tier while rebuilding the Pro tier from scratch after scrapping the original architecture. Whenever Gemini 4 arrives, it will land into a market where OpenAI’s GPT-5.6 and Anthropic’s Fable 5 have had months to establish themselves.
That’s your AI news for Thursday, July 24. Big open-weights moment coming Sunday with Kimi K3 — we’ll have full coverage. Drop a comment below with your take on the EU AI Act crunch.