Top Stories
Google Launches Two New AI Chips to Challenge Nvidia
Google unveiled its 8th-generation TPU processors at Cloud Next 2026 on April 22, splitting its chip strategy into two specialized designs for the first time: the TPU 8t for model training and TPU 8i for inference.
The training chip delivers 2.8 times the performance of the 7th-generation Ironwood TPU at the same price. The inference chip takes a different approach, packing 384 megabytes of SRAM per chip — triple what Ironwood offered — to speed up the fast-response workloads that power chatbots and API calls.
Google isn’t replacing Nvidia here. It still buys plenty of Nvidia hardware. But these in-house chips give Google Cloud customers an alternative, and adoption is growing: all 17 U.S. Department of Energy national laboratories now use AI software built on Google’s TPU stack, and Anthropic has committed to using multiple gigawatts worth of TPU capacity. Both chips will become available later this year.
Sources: TechCrunch · CNBC
PwC Study: 20% of Companies Capture 74% of AI’s Economic Value
A new PwC study of 1,217 senior executives across 25 sectors found a widening gap between companies that generate real returns from AI and everyone else. The top 20% of firms capture nearly three-quarters of all AI-driven economic value, delivering 7.2 times the AI-driven financial performance of their peers.
The gap isn’t about who has the most AI projects running. Leading firms are 2.6 times more likely to use AI for business model reinvention rather than just cost-cutting. They’re also nearly twice as likely to let AI execute multiple tasks within guardrails (1.8x) and operate in autonomous, self-optimizing ways (1.9x). Perhaps most telling: these companies are increasing the number of decisions made without human intervention at 2.8 times the rate of their peers.
The study also found a trust divide. Employees at top-performing companies are twice as likely to trust AI outputs, and those firms are 1.7 times more likely to have a responsible AI framework in place.
Source: PwC
Connecticut Senate Passes Sweeping AI Regulation Bill
The Connecticut Senate voted 32-4 to pass an amended Senate Bill 5, one of the most comprehensive state-level AI proposals in the country. The bill regulates developers of “frontier” AI models, creates a state AI sandbox for companies to test new technologies, and introduces new rules around youth social media and AI chatbot use.
One notable provision: AI chatbot operators would need to make reasonable efforts to detect suicidal ideation or signs of self-harm from users and respond with appropriate resources. The bill cites data showing over 70% of teenagers use AI companions, with about half using them regularly — and documented cases where chatbots failed to intervene or even encouraged self-harm in minors.
The bill also requires employers to notify workers when AI is used in hiring and employment decisions, and bars discriminatory use of AI decision-making tools. It now heads to the Connecticut House, which declined to act on last year’s AI measure.
Source: CT Mirror
Quick Hits
-
MIT Tech Review launches “10 Things That Matter in AI”: The new annual list replaces their traditional breakthrough technologies format for AI coverage, highlighting world models, enhanced LLMs, AI agents, military AI, deepfakes, China’s open-source bet, and AI-powered research as the defining trends of 2026. MIT Technology Review
-
Robinhood Ventures invests $75M in OpenAI: The Robinhood Ventures Fund I purchased approximately $75 million of common stock on April 17, joining the $122 billion funding round that valued OpenAI at $852 billion. GlobeNewsWire
-
Tech layoffs surpass 81,000 in 2026: With nearly half the cuts attributed directly to AI-driven restructuring, Oracle (30,000 jobs), Snap (1,000 jobs), and Meta lead the reductions. Snap CEO Evan Spiegel cited AI automation as enabling over $500 million in savings by H2 2026. TechRepublic · NewsBytesApp
-
Qwen 3.6-Max-Preview tops coding benchmarks: Alibaba’s latest proprietary model, released April 20, now ranks first on SWE-Bench Pro, Terminal-Bench 2.0, and four other major coding evaluations — marking the first time a Chinese lab’s model has held the top spot across so many coding benchmarks simultaneously. Qwen Research
Worth Watching
The AI chip race is heating up. Google’s split into dedicated training and inference TPUs mirrors a broader industry trend: the era of one-size-fits-all AI hardware is ending. With Amazon building its own Trainium chips, Microsoft developing Maia, and Nvidia dominating the high end, the question is whether specialized chips will actually bring inference costs down enough to shift the market — or whether Nvidia’s software ecosystem keeps it entrenched. Watch for pricing details when Google’s 8th-gen TPUs launch later this year.
State AI regulation is moving fast. Connecticut joins New York (with the RAISE Act effective January 2027) and Colorado (AI Act effective June 2026) in passing substantive AI legislation. With over 300 AI bills being tracked across U.S. states and the White House pushing for federal preemption of state laws, the patchwork is growing faster than any federal framework can standardize it. Companies building AI products need to pay attention now, not later.