Plumbing the
information age

Fabric.AI and Kopin are on track for a working CES 2027 demo (January 2027) after advancing through engineering and integration milestones since May 2026. Neural I/o uses MicroLED pixels as optical transceivers for GPU-to-GPU, board-to-board, and rack-to-rack communications, eliminating copper wiring and expensive laser systems. The architecture promises sub-picojoule-per-bit efficiency—orders of magnitude lower power than conventional interconnects, with direct implications for data center power budgets and cooling. Fabric.AI holds ~$30M in cash and has NDAs with unnamed chipmakers for integration. As GPU clusters scale to hundreds of thousands of accelerators, interconnect power and thermal density become limiting factors ahead of total bandwidth. An optical approach with sub-picojoule efficiency could reshape TCO models for frontier AI deployments. However, the January 2027 demo is still months away; first production units likely 2027-2028. This represents an alternative pathway alongside NVIDIA's silicon photonics (Spectrum-X Photonics) and traditional pluggable optics.

Kavukcuoglu, previously DeepMind's CTO and Google's chief AI architect, inherits leadership of the organization at a critical inflection point. Google hasn't released a frontier model since early 2026, while OpenAI and Anthropic have shipped advanced systems in recent months that have demonstrated superior capabilities in coding and complex reasoning tasks. Kavukcuoglu's appointment signals a strategic shift toward LLM improvements as the primary route forward, abandoning the previous emphasis on AGI research beyond language models. Analyst commentary suggests this reorganization better positions Google in the active frontier model race. The move signals renewed engineering focus on model capabilities rather than broader AGI ambitions, directly addressing the competitive pressure from rivals shipping more frequently and performing strongly on benchmark reasoning tasks.

DeepSeek released V4-Pro after the model had been in preview since April. The model handles a context window of up to 1 million tokens and can produce outputs as long as 384,000 tokens, with the option to run in either thinking or non-thinking mode. V4 Pro is now at $0.435/$0.87 per 1M tokens in/out as the published list price, roughly 35× cheaper on input and 86× cheaper on output than Claude Opus 4.7. However, on Terminal-Bench using the reference Terminus 2 harness, V4-Pro scores 54.68% while its cheaper sibling V4-Flash scores 67.04% — a 33-point gap that represents a categorical outlier. A price increase is set to take effect at 16:00 UTC on August 16, with DeepSeek introducing peak and off-peak billing at half the peak-hour price. This GA launch resolves the competitive positioning question after DeepSeek's cheaper V4 Flash model unexpectedly outperformed the April preview of V4 Pro in several independent tests.

xAI shipped Grok 4.6 only 35 days after Grok 4.5, with the underlying checkpoint dated August 10 and a 500,000 token context window. The model builds on previous iterations with particular focus on long-running agents and more ambitious interactive and visual work, staying with complex tasks across many steps whether researching a topic, analyzing information, working across a codebase, or turning an idea into a polished application. However, on Terminal-Bench v3.0, Grok 4.6 scores just 26%, far below the 34.6% of GPT-5.6 Sol Max—a significant gap for workloads like DevOps agents executing infrastructure commands autonomously. xAI's announcement emphasizes Grok Build, Cursor, and the API, indicating this is a developer-first release rather than a consumer product refresh. According to industry tracking, Grok 4.7 is expected within weeks and Grok 5 is targeted before the end of 2026, suggesting an aggressive release cadence.

Databricks closed a $5 billion strategic financing round at $190 billion valuation, led by Coatue Management alongside Blackstone, MGX, and T. Rowe Price entities. This marks a steep valuation climb from $134 billion six months earlier, signaling investor confidence in the enterprise AI agent infrastructure space. The San Francisco company crossed a $7 billion annualized revenue run rate in its second quarter with more than 80% year-over-year growth. Lakebase, a serverless Postgres database for AI-agent workloads, has surpassed a $100 million revenue run-rate. CEO Ali Ghodsi noted that skyrocketing AI costs are boosting demand for the company's AI Gateway platform and open-source tools, with many customers increasingly adopting Chinese open-weight models despite prior hesitations. For infrastructure practitioners, this underscores market bifurcation around AI agent deployment and growing importance of multi-model cost optimization strategies.

Following formal execution of a definitive agreement in early August 2026, SpaceX is progressing toward final closing of its $60 billion all-stock acquisition of Anysphere, developer of the AI-powered coding assistant Cursor. Cursor is used by more than 50,000 businesses and over 64% of Fortune 500 companies, providing SpaceX's AI division with an established enterprise customer base and substantial coding-interaction data. SpaceX's second-quarter 2026 report confirmed revenues reached $7.8 billion—a 92% year-over-year increase—with expansion driven by enterprise cloud agreements and Starlink growth. Anysphere held $4 billion in annual recurring revenue before acquisition. Cursor staff were informed at an all-hands meeting that the brand may be phased out, with future products potentially adopting the Grok name. This deal signals consolidation in the AI coding tool space and provides SpaceX direct access to enterprise developer workflows alongside compute infrastructure from its xAI Colossus cluster—operationally significant for vertical integration of software and infrastructure.