Daily Tech Trends Digest

Big Tech's AI capex revolt: Alphabet down ~7%, Meta down ~9.8%, with Microsoft and Meta reporting tomorrow (Jul 28–29). Alphabet's Q2 2026 results (Jul 22) delivered 82% cloud-revenue growth but a record $44.9B quarter of capex and the company's first negative free cash flow (−$5.9B) since its 2004 IPO, sending GOOGL down roughly 7% after hours. Meta shares have since shed ~9.8% ahead of their own earnings, as investor sentiment completes a historic inversion: capex restraint is now rewarded over capex ambition. Microsoft and Meta report Wednesday July 29; Apple and Amazon follow Thursday July 30. The quarter is the first true reckoning for the $700B+ hyperscaler AI infrastructure bet. ^{[12][13][14][15]}

Black Forest Labs launches FLUX 3 — first multimodal model spanning image, video, audio, and robotics from a single backbone (Jul 23). The only major generative AI company still focused on open or semi-open image models has now crossed into omni-modal territory, unifying video clip generation, synchronized audio, and factory-floor robot action prediction in one architecture. FLUX 3 Video and FLUX 3 Action are live via gated partner access; a public open-weight FLUX 3 Dev variant is planned for later in 2026. No public API or pricing has been announced. ^[1][2]

Google drops three Gemini models but delays its flagship 3.5 Pro, while internal morale issues surface (Jul 21–23). Gemini 3.6 Flash landed July 21 at $1.50/$7.50 per million tokens with 17% fewer output tokens than its predecessor and a 280 tok/s throughput, but the awaited Gemini 3.5 Pro — a full architectural rebuild — remains absent. An Axios report citing current and former DeepMind employees attributes the delay partly to morale challenges following key departures, including Gemini co-lead Noam Shazeer (now at OpenAI) and Nobel laureate John Jumper (now at Anthropic). Meanwhile, Google teased Gemini 4 in the same launch event. ^[3][4][5]

Poolside releases Laguna S 2.1 — an open-weight agentic coding model that beats systems 10× its size (Jul 21). The 118B-parameter MoE (8B active per token) scores at the top of SWE-Bench Multilingual and runs a 1M-token context window; its OpenMDW-1.1 license permits commercial self-hosting. Two open-weight frontier coding models (Laguna S 2.1 and Kimi K3) now exist simultaneously — a new baseline for any enterprise willing to run its own inference. ^[6]

EU AI Act's high-risk-system obligations are now 5 days away (August 2). Annex III conformity assessments, CE-marking requirements, and AI Office enforcement powers all become legally binding on August 2, yet surveys cited last week suggest 78% of organisations have not taken meaningful preparatory steps. Any company whose AI-generated output reaches EU residents — regardless of HQ location — is in scope. ^[14]