Today on The Inference Desk, we're tracking a massive new entrant in the open-weight model space: Inkling, a 975B-parameter multimodal system from Mira Murati's Thinking Machines. At the same time, the sheer unit economics of AI adoption are pushing Chinese open models to the top of global download charts. On the enterprise side, major labs are taking integration into their own hands, with Anthropic and Blackstone launching a $1.5 billion applied engineering firm to physically embed AI developers inside customer organizations.
Oracle and AWS both announced new platforms for managing enterprise agents. On Tuesday, Oracle unveiled its AI Agent Studio for Fusion Cloud, providing a developer-native experience with VS Code and Git integration inside a governed framework to combat 'shadow AI'. On Wednesday, AWS and Anthropic announced the Claude Apps Gateway, a self-hosted control plane for managing access, cost, and policy for Claude models. Both systems focus on providing enterprise-grade identity, audit, and policy enforcement for agentic workflows.
Why it matters
The simultaneous launch of these control planes from two cloud giants validates a critical market need: enterprises require robust governance, not just powerful models, to deploy agents in production. For an agentic AI engineer, these platforms represent the emerging best practice for ensuring reliability and security. They shift the focus from ad-hoc prompt engineering to building auditable, policy-driven agentic systems that can safely interact with core business applications, a prerequisite for any real-world enterprise adoption.
A widely-circulated engineering analysis from Wednesday argues that product teams are making a strategic error by focusing on AI chat UIs. Instead, it posits that the critical work is in building 'agent control planes' with robust policy, observability, evaluation, and rollback capabilities. The piece contends that agentic systems that can take action require the same governance rigor as payment systems.
Why it matters
This piece is a clear articulation of the engineering reality for building durable agentic products, directly aligning with your focus as an EIR. It correctly identifies that defensibility and commercial viability will come from the harness—the control plane and its governance features—not the underlying model. This framework is essential for reasoning about reliability, error recovery, and the unit economics of agent products, where unmanaged, non-deterministic behavior can quickly become a catastrophic cost center.
Thinking Machines, a startup founded by ex-OpenAI CTO Mira Murati, on Wednesday released Inkling, a 975-billion-parameter multimodal Mixture-of-Experts (MoE) model. Released under a permissive Apache 2.0 license, the model accepts text, image, and audio inputs and is designed for enterprise agentic AI workloads, with 41B active parameters and a 1-million-token context window. A key feature is 'controllable thinking effort,' a mechanism to balance performance and cost during inference.
Why it matters
The release of a massive, natively multimodal model under a truly open-source license from a high-profile Western lab provides a significant new foundational block for building production agent systems. For an EIR, Inkling represents a potent, customizable alternative to both proprietary models and the growing suite of Chinese open-weight models. Its 'controllable thinking effort' feature directly addresses the cost-performance trade-offs inherent in ML infra and cloud cost optimization, while its open license offers a path to building defensible, specialized agents without dependency on a single provider's API.
Building on the enterprise migration to Chinese open-weight models we've been tracking, a new 'State of Open Source AI' report shows these models now account for 41% of global downloads on Hugging Face—reportedly surpassing US models for the first time. The surge continues to be driven by unit economics; pushing beyond the 30% weekly OpenRouter share we noted earlier this month, startups like Lindy.ai are publicly joining companies like Coinbase in switching production workloads entirely to models like DeepSeek-V4 to slash API bills.
Why it matters
This marks a definitive rebalancing of the open-source ecosystem, cementing the cost-optimization trend we've covered. For engineers, the viable set of production-ready models is increasingly international, with pure cost-performance outweighing geographic origin. While these models often lag the absolute frontier, their price point makes them the default choice for high-volume enterprise tasks.
OpenAI on Wednesday detailed GPT-Red, an automated red-teaming model trained using self-play reinforcement learning to discover and patch vulnerabilities in production models. The company states GPT-Red was instrumental in hardening GPT-5.6 Sol against prompt injection attacks, demonstrating a scalable self-improvement loop for AI safety.
Why it matters
This is a significant application of reinforcement learning for agent safety and reliability. Instead of relying solely on human red-teamers, OpenAI is using RL to create an automated, adversarial training partner that improves model robustness at scale. For engineers building production agents, this points to a future where RL is not just for task capability but is a core part of the security and alignment workflow, making agents more resilient to real-world adversarial inputs.
NVIDIA is promoting 'tokens per watt' as the new critical metric for AI infrastructure, arguing that data center revenue is becoming constrained by power availability, not just raw compute. At an event on Tuesday, the company highlighted its GB300 NVL72 rack-scale system's efficiency gains, alongside software optimizations like disaggregated serving, as essential for maximizing compute within a fixed power budget.
Why it matters
This reframing of performance from FLOPs to power efficiency is a crucial insight for cloud cost engineering. As power and cooling become the primary bottlenecks for data centers, the cost of inference will be increasingly tied to energy consumption. For those managing cloud bills, this means that selecting hardware and software (like vLLM or disaggregated serving architectures) based on their 'tokens per watt' efficiency will become a primary tactic for cost reduction, impacting both GPU selection and deployment architecture.
Anthropic, in partnership with Blackstone and other investors, on Wednesday officially launched Ode, a $1.5 billion enterprise AI services firm. Building on Anthropic's acquisition of Fractional AI, Ode will embed 'Applied AI engineers' within customer organizations—primarily mid-sized businesses in Blackstone's portfolio—to build and implement 'Claude-first' solutions and agentic systems.
Why it matters
This move signals a significant strategic shift for major AI labs, acknowledging that the primary barrier to enterprise adoption is no longer model capability but the complex, hands-on work of integration. As an EIR, Ode's model validates the 'forward-deployed engineering' strategy as a key to commercial traction and defensibility. It suggests that building a successful agentic AI company may require a heavy services component to bridge the gap between powerful technology and messy enterprise reality, directly challenging traditional IT consulting firms.
Insilico Medicine and Bora Pharmaceuticals have formed a strategic alliance, potentially valued at over $2.5 billion, to integrate AI into the full drug lifecycle, from discovery to manufacturing. The partnership will extend Insilico's generative AI platforms, PandaOmics and Chemistry42, beyond R&D to optimize pharmaceutical manufacturing processes, quality systems, and supply chains.
Why it matters
This partnership marks a significant expansion of AI's scope in pharma, moving from the now-common application in drug discovery to the complex, regulated world of manufacturing. For bio-ML, this is a crucial test of whether AI can solve tangible, physical-world problems like process optimization and quality control at scale. If successful, it could create a new standard where AI competency in manufacturing becomes a key differentiator for contract development and manufacturing organizations (CDMOs).
At the Google I/O Connect event in Bengaluru on Wednesday, Karnataka Chief Minister DK Shivakumar announced plans to establish India's first government-backed AI University. The initiative is part of a broader push to make Karnataka an 'AI-native state,' which also includes an AI Innovation Hub, green data centers, and a comprehensive Data Centre Policy to support the required infrastructure. The government also plans to introduce AI education in schools starting from the 6th grade.
Why it matters
This is a significant state-level investment that aims to create an entire talent and research pipeline for AI in India's primary tech hub. For an EIR considering the Indian ecosystem, this signals strong institutional support and a long-term commitment to building the foundational pillars for AI development in Bengaluru. The establishment of a dedicated AI university could create a powerful local nexus of talent, research, and startups.
Following yesterday's news that the Indian government directly tasked BharatGen and Sarvam AI with building sovereign cyber-defense systems, the two domestic firms are now releasing foundation models at aggressively subsidized rates. BharatGen's Param-2 is priced at ₹5 per million output tokens, while Sarvam AI's models are priced at ₹10 and ₹16. The move aims to undercut global competitors and accelerate domestic AI adoption, particularly for Indian language workloads.
Why it matters
This aggressive pricing by domestic firms, enabled by government GPU subsidies, is a clear move to capture the Indian market and reduce reliance on foreign models. For an EIR in the Indian ecosystem, this dramatically changes the unit economics for building local AI products. While these models may not yet match frontier performance, their cost-effectiveness for Indic languages creates a significant market opportunity and a new set of infrastructure trade-offs to consider when building agentic systems for India.
At its Google I/O Connect event in Bengaluru on Wednesday, Google announced a major expansion of its AI ecosystem in India. Key initiatives include making Gemini available on Google Distributed Cloud for sovereign deployments, launching a free 'AI Research Foundations' curriculum with IISc and NASSCOM, and partnering with AIIMS on India-specific MedGemma health models. Gemini Live will also support over 25 Indian languages.
Why it matters
Google's strategy in India is moving beyond simple market access to building foundational infrastructure and talent. For the Indian AI ecosystem, this provides critical tools for both enterprise and research, from sovereign cloud options to specialized health models and free LLM training. For an EIR, this signals a maturing market with increasing access to sophisticated, localized tools and a growing pool of trained developers.
A consensus is forming across several engineering analyses this week: the majority of so-called 'hallucinations' in RAG systems are actually retrieval failures, not generation errors. The articles argue that developers often focus on tweaking LLM prompts when the root cause is flawed context from poor chunking, naive semantic search, or a vocabulary mismatch between the query and the source documents.
Why it matters
This perspective is a crucial corrective for engineers building production RAG systems. It shifts focus from the LLM as the source of all problems to the data engineering pipeline that feeds it. The key takeaway is that to build reliable retrieval systems, one must rigorously measure and optimize the retrieval step independently. For RAG-based agents, this means prioritizing document parsing, chunking strategies, and hybrid search over prompt engineering for the biggest gains in accuracy and trustworthiness.
Open-Weight Model Landscape Reshapes Around Cost and Multimodality The release of Thinking Machines' 975B-parameter multimodal model 'Inkling' adds a major, permissively licensed option for developers. This comes as new data indicates Chinese open models have surpassed 41% of global downloads on Hugging Face, driven by aggressive cost-cutting strategies from US startups. The market is bifurcating between frontier capability and 'good-enough' cost efficiency.
Major AI Labs Launch Billion-Dollar Services Arms to Tackle Integration Anthropic, Blackstone, and other investors have launched 'Ode,' a $1.5B enterprise AI services firm to embed engineers with customers. This follows a similar pattern from OpenAI and Microsoft, signaling that the primary bottleneck for enterprise AI has shifted from model capability to the messy work of integration, deployment, and governance within existing corporate IT systems.
Agentic AI Moves from Chat to Governed Control Planes A strong theme is emerging: production agentic systems require robust 'control planes,' not just chat interfaces. New frameworks from Oracle and AWS/Anthropic, along with engineering analyses, all point to the need for policy, observability, and auditability to manage agent behavior at scale, treating them as critical automation infrastructure.
India Accelerates Sovereign AI Push with Infrastructure and Education A wave of announcements from India includes plans for the nation's first government-backed AI University in Karnataka, a ack of funding, and new education programs from Google. Local firms are also offering models at a fraction of global prices, supported by government subsidies. The key question is whether these efforts will translate into a sustainable, competitive ecosystem.
AI in Bio-Pharma Broadens from Discovery to Manufacturing Insilico Medicine has partnered with Bora Pharmaceuticals in a deal potentially worth over $2.5B, aiming to apply AI to the entire drug lifecycle, from discovery to manufacturing. This, along with Latent Labs opening access to its protein design agent and a researcher from OpenAI reportedly leaving to start a new drug discovery firm, shows AI's role expanding from pure research into solving complex production and manufacturing bottlenecks.
What to Expect
2026-07-29—US Federal Reserve FOMC meeting and interest rate decision.
2026-07-30—Earnings reports from Amazon and Apple.
2026-07-31—US Personal Consumption Expenditures (PCE) inflation data for June to be released.
— The Inference Desk
🎙 Listen as a podcast
Subscribe in your favorite podcast app to get each new briefing delivered automatically as audio.
Apple Podcasts
Library tab → ••• menu → Follow a Show by URL → paste