Technology Sector
AI infrastructure is fracturing into a cost-efficiency war while the physical grid struggles to keep up.
Alphabet is rolling out three new Gemini models designed to undercut rivals on price while closing product gaps, including Gemini 3.5 Flash Cyber for cybersecurity work and Gemini 3.6 Flash for general tasks (CNBC Tech). The 3.6 Flash variant uses up to 17% fewer tokens than its predecessor and costs less per task than OpenAI's GPT-5.6 Terra Max and Chinese competitors' models, according to Alphabet. Google is also pretraining Gemini 4 and testing Gemini 3.5 Pro with partners, signaling a broader push to rebuild momentum after delays and the rise of Chinese rivals like Moonshot AI's Kimi K3. The moves reflect a strategic bet that efficiency and cost can offset Google's slower timing in key product categories, even as the company develops specialized chips to run Gemini up to 10 times more efficiently.
Nvidia detailed its custom-designed Vera data center CPU this week, marking an aggressive move to vertically integrate its AI infrastructure and lock in customers beyond GPUs alone (CNBC Tech). The chip delivers 50% better performance for AI agents than x86 processors from Intel and AMD by focusing on single-core speed rather than core count, addressing a bottleneck created by agentic AI workloads that require fast CPU-to-GPU handoffs. Vera has already shipped to OpenAI, Anthropic, and SpaceX in June, with OpenAI planning large-scale deployment starting this quarter, though adoption remains early-stage and cloud providers have deep incumbent ties to AMD and Intel. Wolfe Research estimates Nvidia could ship 1.3 million Vera chips this year at roughly $5,000 per chip, tapping into a server CPU market Nvidia estimates could reach $200 billion—vastly larger than the $37 billion server CPU market of 2025.
Intel CEO Lip-Bu Tan scored a notable win Tuesday with Fortinet agreeing to manufacture its next-generation SP6 security chip on Intel 4 process technology. The deal matters because Intel is still hunting for a "significant" marquee customer to justify billions in capital spending on advanced fabs, and while Fortinet isn't Apple or a hyperscaler, the cybersecurity vendor is riding strong momentum—its stock is up more than 100% this year as AI-driven security demand accelerates. Intel's foundry business remains largely dependent on itself, the U.S. government, and vague announcements around Microsoft, Amazon, and Google; a confirmed commercial customer outside defense and internal use signals Tan's strategy to position Intel as a credible alternative to Taiwan's TSMC. The chipmaker reports Q2 earnings Thursday, and investors will be watching for color on the pipeline of customers for its more advanced 14A and 18A processes, which haven't yet landed any public wins at scale.