Google's Gemini 3.5 Pro slips behind schedule amid benchmark misses and internal turmoil

Google's next flagship AI model is months behind its internal targets, and the ripple effects are landing on enterprise teams that have built roadmaps around Gemini's capabilities. CNBC's MacKenzie Sigalos reported this week that Gemini 3.5 Pro has fallen short of internal performance benchmarks, with coding tasks representing a particular gap. The situation is compounded by a complex internal release process, employee frustration, and departures of AI researchers, even as competitors OpenAI and Anthropic maintain aggressive shipping cadences.

The delay matters beyond Google's own product calendar. Gemini is now embedded in third-party platforms, from enterprise productivity tools to industrial hardware, which means schedule slippage at the model layer translates into uncertainty for any operator whose roadmap depends on a specific capability milestone.

Compute scarcity is reshaping vendor strategy

The Gemini delay is one symptom of a broader infrastructure problem. According to B2B Tech News, Google capped Meta's access to its Gemini compute nodes in July, citing an inability to meet the volume of API requests. The constraint forced Meta's internal automation workflows to scale back processing queues to avoid service outages. The episode illustrates a dynamic that procurement and IT leaders increasingly need to plan around: at the frontier of AI, raw compute availability is now a harder constraint than budget or talent.

Anthropic is responding to this dynamic with a structural fix. The company has entered advanced negotiations with Samsung to co-develop a custom AI chip optimized specifically for the Claude model architecture, according to B2B Tech News. The goal is to reduce dependence on third-party GPU suppliers whose capacity is increasingly rationed across competing enterprise customers. Anthropic's engineering teams are designing custom instruction sets intended to accelerate Claude Code workloads natively inside enterprise data centers, with prototype tapeout verification expected in the coming months. The move positions the company to control its own silicon supply heading into an anticipated public market debut.

For enterprise buyers, the Anthropic-Samsung talks carry a concrete procurement implication. If Claude's inference costs drop materially once proprietary silicon is in production, pricing models for existing Claude enterprise contracts could shift. Teams currently benchmarking cost-per-token across providers should factor hardware ownership into their long-term evaluations, not just today's list prices.

Gemini goes to the factory floor and the 6G test range

Even as Gemini 3.5 Pro's flagship release faces delays, Google's existing models are being deployed in demanding operational environments. Boston Dynamics announced a deep integration with Google Cloud on July 14, embedding the Gemini Robotics-ER 1.6 model directly into its Spot quadruped robot, according to B2B Tech News. The integration allows Spot to process layered natural language commands and detect environmental anomalies in real time, reducing the need for continuous human teleoperation. Beta testing across global facilities is scheduled to expand in the coming weeks.

The Spot deployment is a concrete data point for facilities, manufacturing, and logistics operations teams evaluating autonomous inspection or security patrol use cases. A robot that can receive plain-language instructions from a site manager and flag anomalies without a dedicated operator changes the staffing math for large industrial footprints. The open question is how the platform's performance holds when Gemini model updates are delayed or rolled back, a dependency risk that facility operators should address in any service agreement with Boston Dynamics.

AI security testing moves into 6G infrastructure

On the network security side, VIAVI Solutions secured a $1.1 million research grant from the European Smart Networks and Services Joint Undertaking to join the SHIELD6G consortium, according to B2B Tech News. The initiative, announced July 14, is building AI-driven threat simulation platforms designed to stress-test next-generation cellular infrastructure before commercial deployment. VIAVI's specific role involves simulating hyperscale distributed denial-of-service scenarios across European 6G test environments. Simulation rollouts are slated to begin next quarter.

The consortium's approach, replacing periodic manual penetration testing with continuous machine learning vulnerability scanning, has direct relevance for telecom operators and enterprise network teams building private 5G or early 6G infrastructure. Traditional point-in-time security audits leave long windows of exposure between tests. Automated, ongoing simulation narrows that window considerably, and SHIELD6G's results will likely inform procurement standards for 6G network equipment across Europe.

Frontier AI talent commands extraordinary early capital

The compute and talent dynamics playing out at Google are also reshaping the startup funding environment. TechCrunch reported this week on Andrew Dai, a former Google DeepMind researcher who founded Elorian and raised a $55 million seed round at a $300 million valuation before the company had released a product or generated revenue. Isabelle Johannessen of TechCrunch's Build Mode series spoke with Dai about the fundraise, which ranks among the largest seed rounds of 2026.

The Elorian round is notable for enterprise procurement and vendor evaluation teams for a different reason than investor optics. It confirms that AI researchers departing hyperscalers, a trend CNBC flagged in the context of Google's Gemini delays, are attracting institutional capital at a pace that will produce a new wave of enterprise AI vendors within 12 to 18 months. Teams building two- to three-year AI procurement roadmaps should anticipate a significantly wider vendor field, with products originating from researchers who spent years building the exact models now running in production at Google, Anthropic, and their peers.

The immediate operational signal across all of these developments is consistent: compute infrastructure is the binding constraint in enterprise AI right now, and every major provider, from Google to Anthropic to VIAVI, is reorganizing around that reality. The next milestone to watch is whether Anthropic's Samsung chip talks produce an announced development agreement, which would be the clearest sign yet that AI software companies are prepared to vertically integrate hardware to escape the capacity bottleneck.