Five stories from the last few days that actually matter if you build, secure, or ship AI systems for a living. Tap through the log tabs below for the summary, why it matters, and the source on each one.
[2026-07-22 gov] moonshot-distillation-accusation
What happened: White House OSTP Director Michael Kratsios said Moonshot AI distilled Anthropic’s Fable model to build Kimi K3, calling it large-scale, covert industrial distillation, and separately alleged Moonshot accessed restricted Nvidia GB300 chips through Thailand. This is a new escalation on top of Kimi K3’s launch and sellout coverage from earlier this month, the first time a senior US official has named a specific American model as the distillation source.
Why it matters: If accurate, this reshapes how much you can trust “open weight” claims from any lab, and signals tighter export enforcement and possible legal exposure for teams evaluating Kimi K3 for production use.
Source: Reuters technology coverage
[2026-07-22 infra] openai-750b-compute-forecast
What happened: OpenAI raised its projected compute spend through 2030 to roughly 750 billion, up from 600 billion, alongside a new 20 billion data center campus in Georgia known as Project Camellia. The company reported 2025 losses of 38.5 billion as infrastructure and research costs continue to outpace revenue.
Why it matters: This is the clearest signal yet of how much capital the frontier labs think inference and training will cost through the decade, and it puts more pressure on every team downstream to justify model spend against real ROI, not just capability.
[2026-07-22 product] openai-presence-launch
What happened: OpenAI launched Presence, a managed enterprise platform for building and governing production-grade AI agents across voice and chat, pairing model reasoning with company-defined policies, guardrails, and escalation rules. It launched in limited availability, following ChatGPT Work earlier this month.
Why it matters: This is a direct move from selling raw model access to selling governed agent infrastructure, which is exactly the space independent orchestration tooling and in-house agent platforms now have to compete with.
Source: TechCrunch, AI coverage
[2026-07-22 platform] meta-moderation-false-positives
What happened: Instagram and Facebook users reported that Meta’s automated moderation wrongly deleted accounts built up over years, including one with close to a million followers, with Meta restoring some only after journalists intervened.
Why it matters: Automated moderation at platform scale is itself a production AI system with real failure modes and no easy appeal path, a cautionary case for anyone building automated decisioning that affects real accounts or revenue.
Source: The Verge, AI coverage
[2026-07-21 security] kratos-phishing-takedown
What happened: German and US authorities dismantled Kratos, a phishing-as-a-service platform blamed for around 15,000 campaigns a month across 35 countries, seizing over 200 servers and arresting its alleged developer in Indonesia. Trend Micro’s threat intel tooling assisted the investigation.
Why it matters: Kratos-style kits lowered the skill floor for credential phishing at scale, and this takedown, backed by AI-assisted threat hunting, is a useful reference point for how automated defense is starting to match automated attack tooling.
Source: The Hacker News
Moonshot Faces a Distillation Accusation From the White House
White House OSTP Director Michael Kratsios said on July 22 that Moonshot AI distilled Anthropic's Fable model to build Kimi K3, its 2.8 trillion parameter open weight model that topped the Frontend Code Arena earlier this month. He called it large-scale, covert industrial distillation and separately alleged Moonshot accessed restricted Nvidia GB300 chips through Thailand. This is a new development, not a repeat of earlier Kimi K3 launch coverage: it is the first time a senior US official has named a specific American model as the source of an alleged distillation. For teams evaluating open weight models, it is a reason to look harder at provenance claims before trusting a benchmark score.
OpenAI Raises Its 2030 Compute Forecast to 750 Billion
OpenAI now projects roughly 750 billion in compute spend through 2030, up from an earlier 600 billion figure, and announced a new 20 billion data center campus in Georgia called Project Camellia. The company reported 2025 losses of 38.5 billion as infrastructure and research costs continue to outrun revenue. If the largest lab in the market thinks inference and training costs keep climbing this steeply, it is worth budgeting accordingly on your own side of the API.
OpenAI Launches Presence, a Governed Agent Platform
OpenAI introduced Presence on July 22, a managed platform for deploying production grade AI agents across voice and chat, pairing model reasoning with company defined policies, guardrails, and escalation rules. It shipped in limited availability, following ChatGPT Work's launch earlier in the month. The bigger story is strategic: OpenAI is moving from selling raw model access to selling governed agent infrastructure, competing more directly with the orchestration tooling teams already build in house.
Meta's Automated Moderation Deletes Real Accounts
Instagram and Facebook users reported that Meta's automated moderation wrongly deleted accounts built over years, including one with nearly a million followers, with some restored only after journalists asked questions. It is a reminder that automated moderation is itself a production AI system, with false positive rates that carry real consequences and, for most users, no meaningful appeal path.
Kratos Phishing Platform Dismantled
German and US law enforcement took down Kratos, a phishing-as-a-service kit blamed for roughly 15,000 campaigns a month across 35 countries, seizing more than 200 servers and arresting its alleged developer in Indonesia. Trend Micro's threat intelligence tooling assisted the operation, a useful data point on how AI-assisted threat hunting is starting to keep pace with automated, kit-based attack tooling.
FAQ
Is the Kimi K3 distillation accusation confirmed?
No independent verification has been published as of this writing. It is a public accusation from a senior US official, not a court finding, and worth tracking as it develops rather than treating as settled.
Does OpenAI's compute forecast affect API pricing?
Not directly or immediately, but rising committed infrastructure spend across the frontier labs is one of the underlying pressures that shapes pricing and availability over time.
Want a working session on what any of these mean for your team's stack? Tha-Shed's course catalog covers AI tooling for DevOps, cybersecurity, and engineering teams.
