Commercial data centers become targets & Microsoft and Google launch models - Tech News (Apr 3, 2026)

First up, the most unsettling infrastructure story of the week. Reports say Iranian Shahed drones hit Amazon Web Services data centers in the UAE before dawn on March 1, with additional reported strikes on an AWS facility in Bahrain on April 1, and an alleged hit on an Oracle data center in Dubai on April 2. If confirmed, it would be a landmark moment: commercial data centers treated as wartime targets, not just collateral. Analysts caution it’s unclear whether these specific facilities were hosting sensitive military workloads, but the signal is hard to miss. Cloud computing and AI are now woven into national security—and that makes big, civilian-run server campuses both valuable and vulnerable.

Staying in the cloud for a second, Cloudflare is putting numbers on a problem many site operators have been feeling: automated requests are now roughly a third of its network traffic, and AI crawlers are a big reason. The twist is how differently these bots behave compared to humans. People congregate on popular pages; crawlers sweep everything, including obscure corners, creating churn that makes caching less effective and pushes more work back to origin servers. Cloudflare and academic partners are now advocating “AI-aware” caching approaches, essentially separating the needs of human browsing from machine harvesting so websites don’t have to choose who gets a fast experience.

Now to the AI model race, where Microsoft is signaling it wants more of the stack under its own roof, even while it keeps close ties to OpenAI. Microsoft AI announced three new foundational models: one for speech-to-text, one for voice generation including custom voices, and another it describes as image-based video generation that had been seen earlier in its testing playground. They’re rolling out through Microsoft’s Foundry platform, with some access also via the MAI Playground. The headline here isn’t just new models—it’s strategy. Microsoft is framing these as practical communication tools and emphasizing cost competitiveness, which is a direct shot at the current economics of AI.

And Microsoft won’t have the “multimodal” spotlight to itself. Google DeepMind released Gemma 4, a new family of vision-capable models that also expands into video, and, on smaller variants, audio input. Two things make this notable. First, Google is arguing it’s squeezing more capability out of smaller footprints—important for on-device AI where privacy and offline operation matter. Second, the licensing shift: Gemma 4 is under Apache 2.0, which is a big green light for commercial use, redistribution, and community modification. In other words, Google is trying to make “local AI” easier to adopt without legal friction, and that can accelerate an ecosystem fast.

OpenAI also made a move that’s less about models and more about influence: it acquired TBPN, an online tech talk show known for fast analysis and executive interviews. Deal terms weren’t disclosed, but OpenAI says the staff will bolster marketing and communications while the show continues with editorial independence. That promise will be tested in practice, because an AI company owning a news-y outlet immediately raises the question of credibility: what happens when the company being covered is also the company signing paychecks? Still, it’s a clear signal that OpenAI wants to be part of the daily news cycle, not just the infrastructure underneath it.

Let’s shift to the web, where a quieter but important standards update landed. The W3C’s Web Machine Learning Working Group published an updated Candidate Recommendation Draft for the Web Neural Network API, or WebNN. The plain-English significance: it’s another step toward running more machine learning directly in the browser with hardware acceleration, without every app reinventing its own approach. The group is also trying to avoid turning ML features into a new tracking tool, adding guidance and guardrails around privacy and capability probing. Candidate Recommendation is essentially “ready for serious testing,” so now the spotlight turns to whether multiple browsers can implement it in compatible ways.

Now, a cluster of stories about AI-generated media and the incentives that shape it. In the U.S. and elsewhere, more than 200 child development experts and advocacy groups are urging Google to stop recommending AI-generated videos to minors on YouTube, and to remove synthetic media from YouTube Kids. Their argument is blunt: mass-produced, low-quality content can hijack attention, and labels don’t help preschoolers who can’t read. At the same time, a separate cultural example shows how fast this stuff can spread: AI-generated TikTok “fruit drama” videos—chaotic, anthropomorphic soap operas—are pulling staggering view counts, sometimes alongside troubling themes. The throughline is the same: recommendation engines reward whatever holds attention, and generative tools make it cheap to flood the zone.

Two medical research updates stood out today, both showing how computation is reshaping healthcare—carefully, and with real caveats. First, researchers reported patient-specific “digital twins” of the heart that can help plan catheter ablation for ventricular tachycardia, a dangerous rhythm disorder. In a small Johns Hopkins trial, pre-planning with these models dramatically shortened time in the procedure while hitting the right targets more quickly. It’s early days and the study was small, but if larger trials confirm it, this could mean safer procedures and better odds for patients. Second, Johns Hopkins and collaborators reported preclinical results for small molecules that inhibit two hypoxia-related pathways tumors use to survive and resist immune attack. In mice, pairing those inhibitors with checkpoint immunotherapy produced many complete remissions. It’s not a human treatment yet—but it’s a notable direction for overcoming immunotherapy resistance.

Space, quickly—because it’s been a big week. NASA’s Artemis II has launched, sending a crewed Orion mission toward a lunar flyby. It’s the first crewed mission of the Artemis program and a major validation step before future landings. Beyond the engineering milestone, Artemis is now entangled with geopolitics and rules-setting. Multiple nations want a foothold near the Moon’s south pole, where resources could support longer stays. That puts pressure on space law that bans appropriation, while “safety zones” and commercial extraction ideas push boundaries. Even if everyone says they want cooperation, the incentives look increasingly competitive.

Finally, a robotics milestone that’s small in scale but big in implication. Sanctuary AI released a demo of its robotic hand repeatedly reorienting a cube using only fingertip contact—no palm assist—and doing it reliably across multiple attempts. Why that matters: in-hand manipulation is one of the hardest building blocks for robots that can do useful work in human environments, from handling tools to fitting parts together. Sanctuary also claims the behavior transferred from simulation to the real hand without extra training, which—if it holds up broadly—could speed development by reducing the grind of collecting real-world training data.