May 10 Roundup: OpenAI ships GPT‑Realtime‑2 and an open networking standard, Anthropic locks in SpaceX compute, Google rebuilds AI Search around real sources, and ambient AI gets closer
OpenAI moves voice from a demo into production with GPT‑Realtime‑2 and open-sources MRC, the networking protocol behind Stargate. Anthropic picks up 300 megawatts of SpaceX compute and stands up a new enterprise services firm with Blackstone, Hellman & Friedman, and Goldman Sachs. Google makes Gemini API File Search multimodal and rewires AI Search around forum and subscription sources. The Pentagon hands seven AI vendors classified-network access while keeping Anthropic blacklisted. Mistral launches Forge for proprietary enterprise models. And Peter Diamandis sketches what AI will actually feel like in two years — ambient, anticipatory, and sitting on top of every device you own.
1. OpenAI moves voice from demo to production with GPT‑Realtime‑2
OpenAI shipped three new audio models in the API — GPT‑Realtime‑2, GPT‑Realtime‑Translate, and GPT‑Realtime‑Whisper — and the framing matters as much as the spec sheet. The company is explicitly trying to push voice past “fast turn-taking” into voice agents that can actually do work.
GPT‑Realtime‑2 is OpenAI’s first realtime model with GPT‑5‑class reasoning. It quadruples the realtime context window from 32K to 128K tokens, supports parallel tool calls with audible “tool transparency” phrases (“checking your calendar,” “looking that up now”), adds preambles, recovery behavior, and adjustable reasoning effort from minimal to xhigh. On Big Bench Audio, GPT‑Realtime‑2 (high) scores 15.2% higher than GPT‑Realtime‑1.5; on Audio MultiChallenge, GPT‑Realtime‑2 (xhigh) is 13.8% higher.
The translation model handles 70+ input languages into 13 output languages while keeping pace with the speaker. The new Whisper variant streams transcription as people talk. Pricing: GPT‑Realtime‑2 at $32 per million audio input tokens and $64 per million output, GPT‑Realtime‑Translate at $0.034/minute, GPT‑Realtime‑Whisper at $0.017/minute.
“What stood out about GPT-Realtime-2 was the intelligence and tool-calling reliability it brings to complex voice interactions. On our hardest adversarial benchmark, this translates to a 26-point lift in call success rate after prompt optimization (95% vs. 69%). The combination of agentic competence and guardrail strength is what makes it viable for production voice at Zillow.” — Josh Weisberg, SVP and Head of AI, Zillow
Early customers cited by OpenAI include Zillow, Priceline, Vimeo, Deutsche Telekom, and BolnaAI — pointing at three production patterns: voice-to-action, systems-to-voice, and voice-to-voice. The Realtime API also runs active classifiers over sessions and pairs with the Agents SDK for guardrails.
Voice has been the most over-promised, under-delivered modality of the LLM era. GPT‑Realtime‑2 is the first model where “stick a real agent on the phone” is a believable engineering plan instead of a demo script. The 128K context window plus parallel tool calls and reasoning effort knobs means contact-center, field-ops, and clinical-intake voice agents can finally hold a multi-turn workflow without the dialog manager exploding. Enterprises that have been stalling voice rollouts waiting for “GPT‑5 for audio” should start scoping pilots this quarter — and treat tool transparency phrases as a UX requirement, not a feature.
2. OpenAI open-sources MRC, the networking standard behind Stargate
Less flashy but arguably more strategic: OpenAI released the MRC (Multipath Reliable Connection) specification through the Open Compute Project, co-developed with AMD, Broadcom, Intel, Microsoft, and NVIDIA over the past two years.
MRC extends RDMA over Converged Ethernet with SRv6 source routing. Instead of pinning each transfer to a single path, MRC sprays packets across hundreds of paths and reorders them at the destination. It supports multi-plane networks — splitting an 800Gb/s interface into eight 100Gb/s planes — which means a single 64‑port switch can fan out to 512 ports and connect roughly 131,000 GPUs in just two switching tiers instead of the usual three or four.
OpenAI says MRC is already running on its largest NVIDIA GB200 supercomputers, including the Abilene, Texas Stargate site with Oracle and Microsoft’s Fairwater clusters, and that multiple OpenAI models have been trained on it.
“Our goal was not just to build a fast network, but also to build one that delivers very predictable performance, even in the presence of failures, to keep training jobs moving.” — OpenAI Scaling Team
This is the kind of move that quietly resets industry economics. By open-sourcing the protocol that lets Stargate-class clusters survive link failures and avoid hot spots, OpenAI is making sure the next generation of frontier compute — including everyone else’s — converges on a network design it co-authored. For enterprise architects: even if you’ll never build a 130K‑GPU fabric, MRC is a leading indicator of where AI infra vendors are heading. Expect SmartNICs and switch firmware refreshes from AMD, Broadcom, and NVIDIA to start asking pointed questions about your existing RoCE deployments within 12 months.
3. Anthropic adds 300MW of SpaceX compute and stands up a new enterprise services firm
Anthropic announced a compute partnership with SpaceX that gives it access to all capacity at the Colossus 1 data center: more than 300 megawatts and over 220,000 NVIDIA GPUs coming online within the month. The company simultaneously raised limits across its product line — doubling Claude Code’s five-hour rate limits for Pro, Max, Team, and seat-based Enterprise; removing peak-hours throttling for Pro and Max; and significantly raising API rate limits on Claude Opus models.
The SpaceX deal stacks on top of Anthropic’s existing roster of compute commitments: a 5GW Amazon agreement (with ~1GW arriving by end of 2026), a 5GW Google/Broadcom partnership starting in 2027, $30B of Azure capacity through Microsoft and NVIDIA, and a $50B Fluidstack-led American AI infrastructure investment. Anthropic also flagged that it’s exploring orbital AI compute with SpaceX — an idea that would have read as theater a year ago.
On the same day, Anthropic announced a new enterprise AI services company with Blackstone, Hellman & Friedman, and Goldman Sachs as founding partners, backed by General Atlantic, Leonard Green, Apollo Global Management, GIC, and Sequoia Capital. The new firm will deploy Claude into mid-sized banks, manufacturers, and regional health systems — segments the big systems integrators don’t reach efficiently.
“Enterprise demand for Claude is significantly outpacing any single delivery model. This new firm brings additional operating capability to the ecosystem and capital from leading alternative asset managers.” — Krishna Rao, CFO, Anthropic
Two stories, one strategy. Anthropic is solving for the bottleneck most analysts kept pointing at — capacity for Opus at peak — and the bottleneck most enterprise buyers keep complaining about: nobody to actually do the implementation. Pairing 300MW of new GPUs with a Blackstone-backed services firm is essentially a vertically integrated answer to “why can’t I just buy Claude as a transformation program.” For mid-market CFOs and CIOs evaluating frontier model partners, this changes the calculus: Anthropic is now resourced like an OEM-and-integrator combo, not just a model lab.
4. Google rebuilds AI Search around real sources — and makes Gemini File Search multimodal
Google rolled out five updates to AI Mode and AI Overviews that all push in the same direction: more visible links to real human sources. AI responses now include suggested follow-on articles at the end of answers, highlight links from a user’s own news subscriptions, surface quotes from forums and creators with names and community labels, place inline links next to specific bullet points, and show hover previews on desktop.
The strategy is partly defensive — TechCrunch points out that the AI Overviews still hallucinate at scale (a recent NYT analysis found ~9 of 10 answers correct, which at trillions of queries is hundreds of thousands of bad answers a minute) — and partly an admission that for many queries, what users actually want is grounded human perspective, not a confident summary.
On the developer side, Google upgraded the Gemini API File Search tool to natively process images and text together via Gemini Embedding 2, added custom metadata key/value labels for query-time filtering, and introduced page-level citations so an answer can point directly to the originating page in a PDF.
Both updates are about the same thing: trust. On the consumer side, Google is bowing to the reality that an AI answer without a verifiable source isn’t worth much. On the developer side, multimodal File Search with page citations turns RAG from a research project into something a compliance officer can sign off on. If you’re building enterprise agents, the right architectural pattern after this week is: ground every claim, cite every page, and surface a one-click jump to the source. Anything less is going to feel obviously broken six months from now.
5. The Pentagon expands classified AI deals — and keeps Anthropic blacklisted
The Department of War struck classified-network agreements with seven AI vendors: OpenAI, Google, Microsoft, Amazon, NVIDIA, xAI (now merged with SpaceX), and the startup Reflection. Anthropic — which previously held a $200M Pentagon deal — was excluded after being designated a supply-chain risk over disagreements about mass surveillance and autonomous-weapons red lines.
Anthropic sued and won a temporary injunction. DoD CTO Emil Michael told CNBC that Anthropic is still considered a supply-chain risk, but called the company’s Mythos cyber-defense model “a separate national security moment” and acknowledged that the NSA is reportedly already using it. Dario Amodei met with senior White House officials in April; President Trump told CNBC a deal is “possible.”
“The Mythos issue that's being dealt with government-wide … is a separate national security moment where we have to make sure that our networks are hardened up, because that model has capabilities that are particular to finding cyber vulnerabilities and patching them.” — Emil Michael, CTO, U.S. Department of War
Federal procurement is now an explicit lever in the AI safety debate. Anthropic is paying a real revenue cost for holding the line on autonomous weapons and mass surveillance; OpenAI, xAI, and Google are happy to negotiate. For regulated enterprises, this matters in two ways. First, model selection is no longer just about benchmarks and price — supplier policy posture is now part of the risk profile. Second, Mythos-class cyber capabilities are quietly going operational inside U.S. networks; expect equivalent capabilities to show up in commercial security tooling within 18 months, and plan defenses accordingly.
6. Mistral launches Forge — proprietary enterprise models, not just fine-tuning APIs
Mistral capped a busy week by launching Forge, a full-cycle enterprise model training platform. Forge supports pretraining on internal datasets, post-training via SFT, DPO, and ODPO, and reinforcement learning pipelines for aligning models to internal policies — packaging the same recipes Mistral’s own scientists use for flagship models.
Crucially, customers can train on their own GPU clusters or on Mistral Compute. When they run on their own infrastructure, Mistral never sees the data. The business model layers a platform license, optional data-pipeline services, and “forward-deployed scientists” — a Palantir-style embedded engineering model — on top.
“AI scientists today are not using fine-tuning APIs. They’re using much more advanced tools, and that’s what Forge is bringing to the table … There’s no platform out there that provides you real-world training recipes that work.” — Elisa Salamanca, Head of Product, Mistral AI
Early Forge customers include ASML, DSO National Laboratories Singapore, the European Space Agency, Ericsson (legacy code translation), a public institution restoring damaged ancient manuscripts, and unnamed hedge funds training models on proprietary quantitative languages.
Forge is the clearest sign yet that the “rent the model” phase of enterprise AI is splitting into two markets. For 80% of use cases, frontier APIs win on cost and capability. For the remaining 20% — defense, finance, regulated healthcare, deep-IP industries — owning the weights and training on private data is becoming a strategic mandate. The right CIO question now isn’t “which model do we standardize on?” but “where in our portfolio is custom-trained worth it, and who is our partner for the IP-sensitive workloads?”
7. Peter Diamandis: what AI will actually feel like in 2028
Peter Diamandis published a long thesis on the experiential shift coming over the next two years, arguing that AI is moving from “something you use” to “a ubiquitous, always-on, ambient intelligence layer that orchestrates your life.” Five concurrent shifts: a personal JARVIS that has access to all your data and proactively takes action; environments (homes, cars, hotels) that adjust to your biometrics in real time; smart-glasses-driven AR that overlays history, translation, shopping, and tutoring on the world; humanoid robots and autonomous vehicles graduating from pilot to consumer; and a continuous-health-data feedback loop that replaces the annual physical with always-on optimization.
“AI stops being a tool you reach for and becomes an invisible layer that anticipates you: making decisions on your behalf that you didn’t even know needed to be made.” — Peter Diamandis
Diamandis ties the shift to concrete signals: Waymo serving millions of rides, Tesla’s Cybercab rollout, Apptronik raising $520M at $5B+ with Google DeepMind, Unitree’s Shanghai IPO filing on a 674% adjusted profit jump, and Meta/Apple smart glasses normalizing the AR display layer.
The Diamandis thesis lines up tightly with this week’s product news. GPT‑Realtime‑2 is the “voice layer” of ambient AI. Google’s grounded AI Search is the “verifiable layer.” Anthropic’s SpaceX compute is the “capacity layer.” Mistral’s Forge is the “sovereign layer.” Enterprises that treat ambient AI as a 2030 problem are going to wake up to consumer expectations they can’t meet — customers who already have a JARVIS at home aren’t going to tolerate an IVR at work. The move now is to architect for ambient: persistent context, proactive triggers, and human-in-the-loop overrides as a first-class UX, not a bolt-on.
Why It All Matters
Today’s pile is unusually coherent. Voice is going production. Networking standards are being open-sourced under the leader’s logo. Compute is being locked in across hyperscalers, satellites, and on-prem clusters. AI Search is being grounded in real sources. Federal customers are picking sides on safety policy with billion-dollar contracts. And the most credible long-range forecast says ambient AI lands inside two years.
For SEN-X clients, the implication is the same it has been all month, with sharper edges: pick your model partners with eyes open on policy and capacity; design every agent for grounded sources and audible tool transparency; assume voice and ambient are part of your customer experience by 2027; and start scoping which workloads need custom-trained models on your own infrastructure rather than rented APIs.
Need help navigating AI for your business?
SEN-X turns developments like these into concrete enterprise strategy — model selection, agent architecture, voice and ambient UX, and sovereign-AI deployment plans.
Contact SEN-X →