The drive for compute sovereignty in China has crossed a major threshold. After DeepSeek launched its custom silicon efforts earlier this week, peers like Zhipu AI are now following suit to bypass US export controls entirely. On the infrastructure front, the very AI gateways enterprises are deploying to manage budgets are becoming active targets for cloud resource hijacking.
Darktrace reported on an incident from Thursday where a publicly exposed LiteLLM-Proxy AI gateway, connected to Amazon Bedrock, was compromised via an open SSH service and used to run cryptomining malware. The attackers were able to leverage the gateway's IAM permissions to access cloud resources.
Why it matters
This incident is a concrete example of how AI gateways, by centralizing model access and cloud credentials, are becoming high-value targets and single points of failure. It serves as a critical warning for organizations to secure their AI control planes with the same rigor as any other privileged infrastructure, as a compromise could grant attackers broad access to sensitive models, data, and underlying cloud services.
SAP has announced updated API policies that will require AI systems to access its enterprise data through approved channels, primarily its new Joule Agent Gateway. The move is intended to standardize access, enhance security, and prevent data misuse, but it will likely force companies with existing direct API integrations to re-architect their AI workflows.
Why it matters
This is a significant move by a major enterprise software vendor to assert control over its data ecosystem in the age of AI. By mandating access through a specific gateway, SAP is positioning itself as a central gatekeeper for AI interactions with its critical business data, a strategy that could increase licensing costs and architectural complexity for its customers while highlighting the growing trend of vendor-specific AI control planes.
In a detailed case study published Friday, KTern.AI described how it used Amazon Bedrock AgentCore and the Strands Agents SDK to build an enterprise-scale agentic AI platform for SAP digital transformation projects. The company reports that using the managed AWS infrastructure enabled them to deploy production agents in 4-6 hours, significantly reducing project timelines and infrastructure costs.
Why it matters
This case study provides a concrete example of an enterprise successfully deploying agentic AI at scale using a managed platform. For those evaluating AI infrastructure, it highlights the practical benefits of services like Bedrock AgentCore for handling complex requirements such as persistent context, secure tool integration, and scalable agent orchestration. It serves as a strong signal for the production-readiness of Amazon's agent-building stack.
Following up on Meta's entry into the paid API market yesterday, early evaluations of Muse Spark 1.1 show a split in performance. While the aggressively priced model leads in tool-use benchmarks for its 1-million-token context window, early tests indicate it currently trails established competitors on standard coding evaluations.
Why it matters
Meta's entry into the paid API market directly challenges the pricing structures of OpenAI and Anthropic, intensifying the ongoing price war. For AI gateways and platforms, Muse Spark 1.1 is now a critical model to support, offering a cost-effective option for developers focused on building complex AI agents. Its availability may force other model providers to adjust their own pricing and feature sets, particularly around agentic capabilities.
Responding to enterprise concerns over unpredictable AI costs, IBM has updated its 'Bob' agentic development platform with a cost analytics dashboard called 'Bobalytics'. Announced Thursday, the update also introduces multi-agent orchestration capabilities, allowing for more granular control and financial visibility over complex AI workflows within the software development lifecycle.
Why it matters
This update from IBM directly targets a major barrier to enterprise AI adoption: the 'token-maxxing' cost crisis. By providing a native observability and cost-control layer, Bobalytics represents a concrete solution for managing the financial governance of agentic AI. This focus on fiscal accountability is a key differentiator in the crowded developer tool market and a critical feature for production platforms.
GitHub is putting OpenAI's newly available tiered GPT-5.6 family to immediate use within Copilot. Enterprise administrators can now set specific routing rules across the Sol, Terra, and Luna models, matching daily coding tasks to the appropriate tier based on complexity and cost.
Why it matters
This move transforms frontier coding models from a monolithic service into a governable platform choice for enterprises. It provides a clear mechanism for managing the cost-performance trade-off in AI-assisted development, allowing organizations to use the powerful but expensive 'Sol' model for complex reasoning while defaulting to the cheaper 'Terra' or 'Luna' for everyday tasks. This is a key step in making agentic coding economically viable at scale.
A new analysis argues that AI observability has evolved beyond operational monitoring to become a foundational component of product infrastructure. It highlights the advanced capabilities of tools like Arize AI and Confident AI, which now offer features such as custom evaluators for quality control and automated dataset curation from production traffic, enabling teams to wire observability data directly into product decisions.
Why it matters
This conceptual shift is critical for building reliable AI applications. By treating observability as a core product component rather than an afterthought, developers can proactively manage issues like model drift, hallucinations, and latency. For platform providers, integrating these advanced observability features is becoming a key differentiator for enabling customers to build and maintain user trust in their AI products.
SambaNova Systems, a developer of AI chips and full-stack enterprise AI platforms, has completed the first close of a $1 billion Series F funding round, valuing the company at $11 billion. The funding, led by General Atlantic, coincides with news that JPMorgan Chase is deploying SambaNova's hardware for secure, on-premise agentic AI inference workloads.
Why it matters
This massive funding round and major enterprise deployment validate the growing market for specialized, full-stack AI solutions, particularly for regulated industries that prioritize on-premise control over public cloud offerings. It signals strong demand for alternatives to Nvidia's general-purpose GPUs for specific, high-stakes workloads like agentic AI in finance.
Beyond the 8.9 million active users we noted in Ollama's $65 million Series B announcement, the local AI platform disclosed it has reached adoption within 85% of Fortune 500 companies. The new capital is explicitly earmarked to build out a hybrid cloud tier, allowing developers to offload larger models that exceed local hardware constraints.
Why it matters
This funding round, which we noted yesterday, underscores the massive developer demand for tools that bridge local and cloud AI development. Ollama's success demonstrates the market's need for flexible, cost-effective ways to use open-weight models, creating a strong alternative and complement to commercial, API-only platforms. Its growth is a key signal of the maturation of the open-source AI ecosystem.
The in-house silicon strategy we tracked DeepSeek initiating earlier this week is expanding. Zhipu AI is now reportedly developing its own custom inference chips as well, aiming to co-design hardware for its GLM-5.2 model. The joint industry move underscores a push to fully bypass both US export controls and domestic hardware suppliers like Huawei.
Why it matters
With multiple top Chinese labs now pursuing full-stack vertical integration, 'compute sovereignty' is moving from a single company's contingency plan to a broader ecosystem shift. If successful, this co-design approach will structurally lock in the cost advantages Chinese models currently enjoy over Western alternatives.
Perplexity has revealed it is using a post-trained version of the Chinese open-source model GLM 5.2 as the default orchestrator for its agentic systems. The company claims this custom model achieves performance comparable to Anthropic's Claude Opus 4.8 for many tasks, but at approximately one-third of the cost. More complex queries are automatically escalated to the more expensive Opus model.
Why it matters
This is a landmark case study in production cost-optimization using open-weight models. Perplexity's multi-model routing strategy demonstrates how to effectively blend high-cost frontier models with cheaper, 'good enough' alternatives to manage operational expenses without sacrificing quality. This hybrid approach is becoming a crucial competency for any team building scalable AI products and a key value proposition for AI gateways.
The 'tokenmaxxing' budget crisis we've been tracking is prompting direct corporate intervention. Building on the recent KPMG survey data regarding usage-based pricing chaos, a new Inc.com report details that major firms—now explicitly including Nvidia alongside previously reported companies like Uber—are restricting internal employee access to certain AI tools to cap spiraling token costs.
Why it matters
The escalating cost of token-based pricing is emerging as a primary obstacle to widespread enterprise AI adoption. This financial friction is creating a strong market pull for AI gateways with sophisticated cost-routing, caching, and budget controls. It also puts pressure on model providers to offer more transparent and predictable pricing, and it validates strategies that leverage cheaper open-weight models for routine tasks.
China's AI Leaders Pursue Vertical Integration with Custom Chips In response to US export controls, major Chinese AI firms like DeepSeek and Zhipu AI are reportedly developing their own custom inference chips. This strategic shift towards 'compute sovereignty' aims to reduce reliance on both foreign and domestic hardware suppliers, optimize performance for their proprietary models, and control long-term operational costs.
AI Gateways Emerge as Critical Infrastructure and Key Security Targets The AI gateway's role as a central control plane for routing, authentication, and governance is being solidified by new enterprise offerings from Citrix and SAP. However, a recent incident involving a compromised LiteLLM gateway highlights their vulnerability, establishing them as high-value targets for attackers seeking to exploit aggregated cloud and model permissions.
Open-Source AI Drives Enterprise Cost Optimization Strategies Enterprises are increasingly adopting open-source models, particularly from China, to manage spiraling AI costs. Perplexity's use of a fine-tuned GLM 5.2 model to handle most tasks—achieving near-Opus performance at a third of the cost—exemplifies a sophisticated, multi-model routing strategy that is becoming a new standard for production AI.
Funding Pours into AI Infrastructure and Specialized Hardware A wave of significant funding rounds for companies like SambaNova ($1B), Nscale (€1.7B), and Ollama ($65M) demonstrates strong investor confidence in the foundational layers of the AI stack. The capital is flowing into specialized inference hardware, decentralized platforms, and developer tools, signaling a market focus on building out the production-ready infrastructure needed to support enterprise-scale AI.
Enterprises Grapple with Spiraling Costs and Vendor Lock-in As AI moves from pilot to production, enterprises are confronting the 'dollar-sign shock' of unpredictable token-based pricing. Reports of companies like Uber limiting AI use and growing CEO concerns over vendor lock-in are driving a market-wide search for better cost controls, multi-model strategies, and transparent governance platforms.
What to Expect
2026-07-17—China's World Artificial Intelligence Conference (WAIC) opens, expected to showcase the country's end-to-end domestic AI stack.
2026-09-01—Meta is scheduled to begin manufacturing its in-house 'Iris' AI chip.
2026-10-20—NVIDIA holds its GTC research keynote, likely detailing further AI infrastructure advancements.
How We Built This Briefing
Every story, researched.
Every story verified across multiple sources before publication.
🔍
Scanned
Across multiple search engines and news databases
523
📖
Read in full
Every article opened, read, and evaluated
196
⭐
Published today
Ranked by importance and verified across sources
12
— The Gateway Signal
🎙 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