🛰️ The Gateway Signal

Wednesday, September 9, 2026

12 stories · Standard format

Generated with AI from public sources. Verify before relying on for decisions.

🎧 Listen to this briefing or subscribe as a podcast →

The profit pools of multi-model orchestration are shifting. With open-source platforms proving they can serve massive reasoning models natively without enterprise gateway tolls, the battle for developers has moved directly into state-aware memory disaggregation and specialized execution hardware.

AI Gateways

EvoLink Adds DeepSeek V4 Flash API Route with Automated Prefix Caching

EvoLink has added the DeepSeek V4 Flash model under the ID `deepseek-v4-flash`, supporting a 1M-token context window, 38K maximum output tokens, optional thinking mode, and automatic prefix caching. The route is accessible across OpenAI Chat Completions, Anthropic Messages, and Responses protocols. Input pricing is set at $0.442 per 1M tokens ($0.015 per 1M for cache hits) with output pricing at $1.324 per 1M tokens, including server-side web search tools billed per successful call.

For your tracking of EvoLink against OpenRouter, Portkey, and LiteLLM, this release shows EvoLink aggressively targeting high-concurrency reasoning workloads with a 96% discount on cached prompt reads. By abstracting DeepSeek's native interface behind unified OpenAI and Anthropic protocol adapters under a single API key, EvoLink lets developers swap underlying reasoning backends without changing client-side SDK logic. The addition of server-side web search tools billed per execution rather than per token also simplifies cost accounting for autonomous research agents.

Verified across 1 sources: EvoLink

Vercel AI Gateway Introduces Team-Wide Zero Data Retention Controls

Vercel AI Gateway launched Zero Data Retention (ZDR) enforcement for Pro and Enterprise plans, toggled team-wide via dashboard settings ($0.10 per 1,000 requests) or per-request using `zeroDataRetention: true`. The feature guarantees immediate deletion of prompts and outputs across supported providers, including Anthropic, Google Vertex AI, AWS Bedrock, Together AI, and Mistral, while introducing granular Bring Your Own Key (BYOK) compliance tracking.

Managed gateways are shifting competition from raw model routing to strict enterprise governance and privacy controls. Adding verified zero-retention guarantees across disparate backends removes a major procurement roadblock for enterprises evaluating multi-model gateway architectures. By charging a flat per-thousand-request governance fee rather than marking up underlying token rates, Vercel establishes a clear value-add model against self-hosted proxies like LiteLLM.

Verified across 1 sources: Vercel

AISIX AI Gateway Adds Context-Preserving Routing for Asynchronous Batch APIs

AISIX AI Gateway introduced native traffic management for OpenAI-compatible Files, Batch, and Fine-tuning APIs. To preserve routing context across asynchronous requests that lack model fields—such as `GET /v1/batches/{id}`—the gateway embeds state context into opaque IDs prefixed with `aisix-`. The system decouples zero-token lifecycle management calls from completed batch usage, enforcing idempotent token accounting and credential isolation across long-running background jobs.

Traditional AI gateways built strictly for synchronous request-response pairs fail when handling asynchronous offline inference and model fine-tuning loops. Embedding state context into transaction IDs allows infrastructure teams to enforce uniform security policies, budget limits, and audit trails across both real-time chat completions and multi-hour batch runs without requiring application code to track endpoint state.

Verified across 1 sources: API7.ai

LLM Inference Platforms

Nvidia Groq 3 LPX Enters Production Delivering 3,400 tok/s for Agent Decode

Moving from the Hot Chips architectural preview we tracked late last month into full production, Nvidia's liquid-cooled Groq 3 LPX rack packages 256 LP30 processors, rack-scale SRAM, and a deterministic compiler to accelerate small-batch decode tasks. Benchmarks by Artificial Analysis recorded output throughput exceeding 3,400 tokens per second on Gemma 4 31B over a 100K context window. Nvidia is pairing LPX with Vera Rubin NVL72 platforms for disaggregated prefill and decode serving, with initial neocloud deployments heading to Nebius.

For platform teams tracking inference economics, Groq 3 LPX demonstrates the transition from uniform GPU clusters to specialized prefill and decode hardware. Long-context agentic interactions are severely constrained by decode latency; utilizing SRAM-centric LPX accelerators specifically for the token-generation phase eliminates tensor-parallelism communication overhead and establishes a benchmark target for competitors like Cerebras Cloud.

Verified across 2 sources: Jon Peddie Research · Narrative News

AI Infrastructure

IBM Research and Red Hat Deploy 753B GLM-5.2 on 544 H100s via llm-d Gateway

Following yesterday's donation of the `llm-d` gateway to the CNCF, IBM Research and Red Hat demonstrated the open-source framework serving the 753-billion-parameter GLM-5.2 model across 544 NVIDIA H100 GPUs. The setup supported up to 3,000 concurrent coding agents with zero preemptions, achieving over 6.6 million output tokens per minute at peak. The architecture leverages prefix-aware routing, peer-to-peer KV-cache sharing, and prefill/decode disaggregation to cut operating costs by 5x to 10x compared to commercial APIs.

This deployment provides concrete empirical data for enterprise teams considering self-hosted infrastructure over commercial gateways like OpenRouter or Together AI. By utilizing `llm-d`'s disaggregated prefill/decode pipeline on existing H100 hardware, platform engineers can serve multi-turn agent loops with heavy context reuse without suffering preemption drops. It demonstrates that open-source Kubernetes-native control planes can achieve hyperscale token throughput while bypassing per-token proxy markups.

Verified across 1 sources: IBM Research Blog

SemiAnalysis Benchmarks Google TPUv7 Ironwood on TorchTPU Stack

Yesterday we covered SemiAnalysis's benchmarks showing Google's TPUv7 Ironwood delivering up to 50% higher throughput per dollar than NVIDIA GPUs on the native TorchTPU stack. Further details from the report confirm Google plans to bring TorchTPU out of private beta in October, expanding support to Kimi K3 and GLM-5.3, utilizing Pallas kernels, StableHLO, and XLA.

The externalization of Google's TPU infrastructure via a native PyTorch interface directly challenges NVIDIA's dominance in commercial hosted inference. By removing custom JAX framework constraints, platform teams can execute standard PyTorch serving pipelines on lower-cost TPU silicon, creating viable alternative hardware targets for high-volume inference providers like Together AI and Fireworks.

Verified across 1 sources: sechub.in

AI Startup Funding

Mistral AI Closes €3B Series D at €21B Valuation for Sovereign AI Infrastructure

French AI lab Mistral AI secured a €3 billion ($3.5 billion) Series D funding round at a post-money valuation exceeding €21 billion, led by Samsung Electronics alongside Scaleup Europe Fund, PSG Equity, ASML, BlackRock, and a16z. The round marks the largest VC financing in European tech history and supports Mistral's shift toward full-stack enterprise sovereign deployments, custom model hosting, and dedicated data center capacity.

Mistral's massive capital infusion highlights an enterprise pivot toward sovereign, self-hosted deployment models where organizations retain total control over compute and weights. Rather than competing purely on raw API token pricing against US hyperscalers, Mistral is building vertically integrated, forward-deployed engineering stacks tailored for highly regulated enterprise environments.

Verified across 2 sources: HPCwire · PitchBook

Fluidstack Secures $1.5B Series C at $18B Valuation for Asset-Light Data Centers

AI data center developer Fluidstack raised $1.5 billion in funding led by Jane Street Capital, bringing its valuation to $18 billion. Unlike neoclouds that purchase massive GPU inventories, Fluidstack focuses on power acquisition, modular facility construction, and cluster management software like Atlas OS. The company is currently building single-tenant facilities for Anthropic in Texas and New York under a $50 billion infrastructure commitment.

Fluidstack's funding structure underscores how capital markets are re-evaluating hardware depreciation risks in AI infrastructure. By operating as a chip-agnostic infrastructure layer that manages physical power and cooling while client labs supply or lease their own silicon, Fluidstack avoids GPU obsolescence while removing the physical facility buildout bottleneck for frontier labs.

Verified across 1 sources: Claypier

China AI Scene

Wuwen Xinqiong and MiniMax Partner to Scale Domestic Chinese Inference Efficiency

Shanghai Wuwen Xinqiong Intelligent Technology and MiniMax signed a strategic partnership to co-optimize LLM inference infrastructure and token scheduling. Wuwen Xinqiong will supply compute power scheduling software and low-level serving optimizations to manage MiniMax's high-volume consumer and enterprise API traffic, targeting operational cost reductions across domestic GPU clusters.

As domestic hardware constraints enforce strict compute budgets across Chinese AI providers, vertical partnerships between specialized scheduling software firms and model labs are critical for operational survival. Coupling software-defined scheduling directly with high-volume token pipelines allows Chinese domestic platforms to optimize token throughput per watt on heterogeneous silicon pools.

Verified across 1 sources: XIX AI

Open Source AI

Stacklok Releases ToolHive for Containerized MCP Server Isolation

Following yesterday's launch of Stacklok's ToolHive platform for MCP server isolation, further technical documentation reveals its containerized runtime natively supports Docker, Podman, and Kubernetes. The release also introduces a semantic tool-search optimizer designed to reduce agent context token usage by up to 85%.

As coding environments like Cursor and Claude Code deploy local MCP servers to connect with internal corporate tools, they expose host systems to credential leakage and unvetted command execution. ToolHive brings containerized zero-trust isolation and SSO governance to agent tool execution, allowing platform administrators to restrict agent access without slowing down developer adoption.

Verified across 2 sources: vuink.com · The Next Gen Tech Insider

vLLM Implements Hybrid HiSparse Offloading for GLM-5.3 Long-Context Serving

The vLLM project detailed Hybrid HiSparse, a memory optimization framework designed to serve GLM-5.3 at a 1-million-token context length on an 8× H200 node. By analyzing sparse-MLA KV-cache activity, Hybrid HiSparse offloads inactive KV tokens to host CPU memory while keeping active pages in GPU memory via a shared Hybrid Memory Allocator, maintaining high decoding concurrency without breaking CUDA graph execution.

Serving 1M-token context windows for complex reasoning models creates severe VRAM bottlenecks that reduce serving batch sizes. Hybrid HiSparse allows open-source serving engines to maintain high query concurrency on existing hardware pools by using system RAM as an extended cache tier, directly lowering per-token costs for long-context agent workloads.

Verified across 1 sources: vLLM

Enterprise AI Adoption

Boomi Launches AI Gateway and Agent Control Plane Built on Lunar.dev

Boomi launched its Agent Control Plane and AI Gateway based on technology acquired from Lunar.dev. Positioned between autonomous agents and enterprise backend systems, the infrastructure layer dynamically intercepts agent traffic, routes routine queries to deterministic software or smaller language models, and enforces centralized cost controls and FinOps attribution.

Uncontrolled agent loops driving up API token spend have forced enterprise IT teams to institute strict traffic governance. Boomi's release reflects a broader trend where traditional integration platforms are deploying inline proxies to inspect agent tool calls, offloading repetitive sub-tasks to deterministic workflows to avoid unnecessary frontier model token expenditure.

Verified across 1 sources: iTWire


The Big Picture

Zero-Markup Gateways Press Commercial Aggregator Margins Open-source proxies and zero-fee platforms like Experiential Labs and Vercel AI Gateway are undercutting commercial providers by routing raw model calls without token markups. By shifting monetization toward enterprise governance features like Zero Data Retention (ZDR) or training telemetry flywheels, these gateways allow development teams to bypass standard aggregator fees while maintaining central control.

Asynchronous Batch and State Management Shift Into the Ingress Layer Gateways are evolving beyond simple synchronous request-response proxies to handle the full lifecycle of long-running AI workloads. Implementations like AISIX embedding routing context into opaque batch IDs illustrate how ingress layers now manage multi-hour fine-tuning, file management, and state polling natively without dropping authorization context.

Heterogeneous Compute Disaggregates Prefill and Decode Execution Driven by the high KV-cache overhead of multi-turn agent sessions, infrastructure architectures are splitting prompt digestion and token generation across distinct hardware pools. Deployments using Groq 3 LPX rack-scale SRAM alongside Vera Rubin GPUs or externalized TPUv7 stacks demonstrate how specialized silicon stages reduce unit token costs.

Containerized Isolation Enforces Zero-Trust Boundaries for Agent MCP Tools As autonomous agents scale tool execution via Model Context Protocol (MCP), open-source security platforms like ToolHive and OpenHands 1.0 are moving tool execution into isolated Docker containers. Placing single-sign-on gateways and risk-analyzing security engines between agents and local environments prevents agents from exposing host credentials.

Sovereign AI Capital Shifts from Raw Silicon Inventory to Power and Infrastructure Operations Megarounds like Fluidstack's $1.5 billion raise and Mistral's €3 billion Series D reflect an infrastructure pivot toward securing gigawatt-scale power, modular site builds, and air-gapped sovereign control. Capital is increasingly backing software-defined infrastructure operators over platforms carrying rapid GPU depreciation risk.

What to Expect

2026-10-01 Google Cloud plans to open source and bring its PyTorch-backed TorchTPU stack out of private beta.
2026-10-01 Cisco and Supermicro begin commercial shipments of integrated Vera Rubin NVL72 liquid-cooled AI factory racks.
2026-12-31 Gartner projects 40% of enterprise applications will deploy task-specific AI agents.

Every story, researched.

Every story verified across multiple sources before publication.

🔍

Scanned

Across multiple search engines and news databases

432
📖

Read in full

Every article opened, read, and evaluated

120

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
Overcast
+ button → Add URL → paste
Pocket Casts
Search bar → paste URL
Castro, AntennaPod, Podcast Addict, Castbox, Podverse, Fountain
Look for Add by URL or paste into search

Spotify isn’t supported yet — it only lists shows from its own directory. Let us know if you need it there.