On September 12, Anthropic CEO Dario Amodei called publicly for independent monitoring of frontier AI models, industry-wide regulation, and a global framework for managing increasingly capable systems. Within hours, OpenAI CEO Sam Altman and Elon Musk expressed support for the direction.[2][6]
This follows Amodei’s June proposal that governments adopt mandatory pre-release testing for frontier AI, drawing an analogy to regulatory approaches used in areas such as pharmaceuticals and aviation. His position was clear: if a model does not meet sufficiently high safety standards, its release should be blocked or reversed.[3]
At Seattle’s Mobile Future Forward conference on September 10, Chetan Sharma interviewed T-Mobile President of Technology and Chief Technology Officer John Saw in a session titled “Engineering the Nervous System of Physical AI.” The program described a future in which intelligence is “dissolving into the physical world — embedded in surfaces, infrastructure, and environments — creating a layer of continuous perception and response that operates below the threshold of human awareness and above the threshold of human control.”[1]
These two parallel conversations materially impact how AI is governed. One speaks to governing AI before it is released. The other is about how to perceive and respond to AI once it is embedded in the physical world.
An important question emerges: How do we regulate AI once it is no longer simply a model, but an evolving system operating in the real world?
That is where the real risk lies. The ability to govern in production relies on the distinction between Human-Time and AI-Time.
Pre-release testing is necessary. It is not enough.
Testing an AI model before release is essential. It gives regulators and developers a point-in-time assessment of what a system can do and whether it meets defined standards.
But the system that is tested before release may not be the system operating months later.
AI-enabled systems can be updated. They can gain access to new tools and data. They can interact with other systems and agents. They can operate in environments that were not present during testing. In some applications, they may learn from real-world use or change through approved update processes.
The FDA already recognizes the reality and the challenges of evolving AI models in medical devices. Its framework for predetermined change control plans allows certain planned modifications to AI-enabled devices to be managed over the product lifecycle, while emphasizing ongoing monitoring, evidence, risk management and the ability to detect and respond to changes in performance.[4]
The European Union’s AI Act similarly extends obligations beyond deployment. High-risk AI providers are required to monitor systems throughout their lifetime, and serious incidents must be reported within specified timeframes.[5][13]
These are important steps, and they acknowledge that production AI needs to be regulated and monitored. However, they still largely operate within a Human-Time regulatory model: something happens, the operator records it, and a regulator reviews the evidence afterward.
The harder question is what happens when the system itself is operating faster than humans can observe, interpret or reconstruct events. This is the AI-Time problem.
One Potential Answer: The Network
John Saw speaks about networks that provide “deterministic connectivity required for real-time control,” where the network stops being “a passive data pipe” and becomes “the active nervous system of physical intelligence itself.”[8]
He describes the transition from informational tokens to what he calls “kinetic tokens” — data that drives action in the physical world, whether through a drone, robotic arm, vehicle or other machine. These applications require ultra-low latency, synchronization across devices and continuous learning at the edge.[7][8]
T-Mobile’s Chief Network Officer, Ankur Kapoor, who also spoke at MFF, said in August that machine-time decisions are already production reality inside the carrier: the network responds to real-time demand and adjusts itself, and “that all happens completely autonomous. No humans involved.”
He named what the network holds that no one else does: “Nothing in the physical AI space is going to work without context … We understand what the context is.”[14]
AT&T sees the same thing from the other side of the industry. Shawn Hakl, SVP of Product for AT&T Business and another MFF speaker, describes the carrier’s custody of an AI transaction end to end: “I have the edge device. I have a RAN (Radio Access Network). I can take that packet, put the deterministic routing around it. I can make it a private connection with policy controllers built in. I can transport that to Nvidia compute that sits at the edge of the network.”[20] His design rule for the agentic era is blunt: “You’ve got to make it agent consumable.”[20] And AT&T’s leadership has put the dependency in one line: “AI doesn’t exist without our connectivity … plain and simple.”[21]
This “layer of continuous perception and response” – the network - could enable AI governance, not only when a model ships, but as it grows and evolves in its real-world applications. That makes the network a candidate for the governance layer. A network is no longer simply carrying information from one place to another. It becomes part of the environment in which AI perceives, decides and acts.
Regulation requires evidence in AI-Time
Humans work in hours, minutes and seconds. The systems increasingly making decisions around us may operate in microseconds or faster. So if those systems are going to be regulated, how do we establish what happened, when it happened, and in what sequence, reliably, at the speed at which the system actually operated?
That means the nervous system Saw describes must do more than simply carry the action. It has to provide the context Kapoor describes and more: timestamp the action contemporaneously, sequence it correctly against every other event it touched, and record it in a form that cannot be altered and that a third party can verify.
Regulators have never operated at the speed of the systems they regulate. The SEC does not stop a trade mid-transaction. It reconstructs the sequence afterward, and it can only do that because the rules force every venue to stamp each event against a common clock. Regulation of fast systems has always been evidentiary rather than real-time, and that evidence has always been based on time.
AI-Time in Financial Markets
Financial markets offer a useful analogy, not because physical AI is identical to electronic trading, but because financial regulators have already confronted the same fundamental problem: you cannot put a human regulator inside every transaction. Instead, the regulatory system creates an evidentiary record, and the regulator reconstructs what happened afterward with a confidence that depends entirely on the quality of that record.
The SEC’s Consolidated Audit Trail requires every reportable event to be captured so it can be linked through the lifecycle of an order, from generation through routing, modification, cancellation and execution. It also sets clock-synchronization standards by participant: exchanges must hold their clocks within 100 microseconds of NIST time, broker-dealers within 50 milliseconds, and manual-order clocks within one second, while firms whose systems produce timestamps finer than a nanosecond must truncate at the nanosecond.[9][10] An order on a US equities exchange today is matched in microseconds.
Spoofing, placing orders you intend to cancel to move the market price, is prosecuted by reconstructing sequence and intent from those records. In 2020 the CFTC (Commodity Futures Trading Commission) ordered JPMorgan to pay $920 million after its traders placed “hundreds of thousands” of futures orders “with the intent to cancel those orders prior to execution.”[11] That case covered eight years of conduct, took years to reconstruct, and the actors were people.
Visa reported in December 2025 that agent-initiated transactions had been completed on its network with more than 100 ecosystem partners and projected that “millions of consumers” would be using AI agents to buy things by the 2026 holiday season.[12]
With all this in mind, now put an AI agent on each side of a trade, negotiating and executing in microseconds across venues whose clocks are allowed to differ by tens of milliseconds. Which order came first? Was the cancel sent before or after the fill? Did the agent see the price it claims it saw? The evidence a regulator would need is an event record that is contemporaneous to the microsecond, sequenced across systems that do not share a clock, unalterable after the fact, and referenced to something the agent’s operator does not control.
And today’s infrastructure does not deliver time across disparate systems at AI-Time precision. The same rulebook that lets firms timestamp in nanoseconds permits two broker-dealers’ clocks to disagree by up to 50 milliseconds, fifty thousand times less precise than the one-microsecond events they are recording.[9][10]
Who establishes the authoritative sequence? And, more importantly, who can independently prove that sequence?
What happens when the regulated system is also the recordkeeper? A log tells us what a system says happened. A regulatory evidentiary layer needs to establish that the record itself can be trusted.
That distinction matters enormously when the entity generating the evidence is also the entity being regulated.
The Industry’s Current Answer: Guardrails, Monitoring and a Human Who Verifies
The technology industry has not ignored what happens after launch. At GTC in March, Nvidia CEO Jensen Huang put the risk plainly: “Agentic systems in the corporate network can access sensitive information, execute code, and communicate externally. Obviously, this can’t possibly be allowed.”[15]
Nvidia’s remedy is a build-deploy-run stack in which run-time guardrails “actively block unsafe model behavior” and “with continuous monitoring … model safety becomes enforceable.” However, by May, Nvidia’s own engineers were writing that scaling agents “with structural transparency and operational integrity requires more than runtime guardrails.”[16]
Qualcomm CEO Cristiano Amon described the operating model for an agent on a device in one sentence: “It plans, it executes, you verify.”[17] Kedar Kondap, Qualcomm’s SVP of Compute and Gaming, expects “hybrid orchestration” of AI across device and cloud to be “where the industry is going”;[18] Ronnie Vasishta, Nvidia’s SVP of Telecom, describes AI-native networks that are “self-adapting.”[19]
Taken together, these perspectives describe a world in which AI action is distributed across the device, edge, network and cloud, and in which the infrastructure itself is continuously changing.
Put simply: guardrails decide whether an action is allowed. Monitoring tells the operator what the system is doing. Verification adds a human check. But none of those, by themselves, answers a regulator, an insurer, a counterparty or any other independent party asking:
What exactly happened? When did it happen? What else was happening at the same time? Was the action appropriate? And how do we know the record is genuine?
Regulating Production AI in AI-Time
None of this is an argument against pre-release regulation. It is an argument that pre-release regulation is just the first half of the problem.
Pre-release testing tells you what the model was capable of on the day it shipped. It tells you nothing about what the model actually did on a Tuesday six months later, at 2:14 p.m. and 37 microseconds, when an agent was negotiating with a system it was never tested against.
This second half of AI governance is not simply post-deployment monitoring. It is a post-deployment regulatory engagement model for AI that evolves and acts in the real world, one that treats every action the system takes as an event that happened at a specific instant, in a specific order, relative to every other event it touched.
That is where governance has to live. Today no one outside the system that acted can prove any of it, and no one can certify, for a regulator or any other third party, an auditable sequence of events across the parties involved.
The next generation of AI regulation therefore requires something we do not yet have at scale: an evidentiary layer capable of operating in AI-Time.
It needs a trusted, compliant and auditable record whose temporal resolution, sequencing and integrity are appropriate to the system being governed, and whose credibility does not depend entirely on the organization that operated the AI.
If AI is going to live in the physical world, regulation needs to meet it there.
Let’s have this conversation now, while the world’s AI leadership is acknowledging the need to monitor, regulate and control AI models.
___________________________________________________________________________
Part of the Ecosystem Entanglement™ series — a practitioner’s view on the physics of ecosystems and the commercialization of deep tech in hyperbolic markets.
Author Lori Salow Marshall is the Co-founder and Managing Member of The Quantum Links™ Group.
Disclosure: Marshall advises companies in the physical AI space who are seeking to solve the regulatory and management challenges of AI-Time vs. Human-Time.
Sources
[1] Mobile Future Forward 2026 agenda and topic descriptions, Seattle, September 10, 2026 — https://www.mobilefutureforward.com/agenda/ and https://www.mobilefutureforward.com/our-topic/
[2] The Washington Post, “Anthropic’s Amodei calls for AI oversight, joined by Altman and Musk,” September 12, 2026 — https://www.washingtonpost.com/technology/2026/09/12/anthropic-ceo-dario-amodei-calls-ai-industry-slow-down/
[3] Axios, “Anthropic CEO says government should block dangerous AI,” June 10, 2026 — https://www.axios.com/2026/06/10/anthropic-ceo-government-block-dangerous-ai
[4] FDA, “Marketing Submission Recommendations for a Predetermined Change Control Plan for Artificial Intelligence-Enabled Device Software Functions,” final guidance, December 4, 2024 — https://www.federalregister.gov/documents/2024/12/04/2024-28361/
[5] Regulation (EU) 2024/1689 (AI Act), Article 72, Post-market monitoring (applies to high-risk systems from August 2, 2026) — https://artificialintelligenceact.eu/article/72/
[6] Axios, “Anthropic, OpenAI CEOs call for slowdown in AI development,” September 12, 2026 (Amodei essay: “We must slow the pace at which we improve the capabilities of AI models”; Altman on X; Musk: “Dario is right”) — https://www.axios.com/2026/09/12/anthropic-ai-amodei-pacing; The Wall Street Journal, “Anthropic Boss Warns AI Industry Must Slow the Pace,” September 12, 2026 — https://www.wsj.com/tech/ai/anthropic-boss-warns-ai-industry-must-slow-the-pace-a4267b56; Bloomberg, “Anthropic CEO Says It’s Time to Slow Pace of Improving AI Models,” September 12, 2026 — https://www.bloomberg.com/news/articles/2026-09-12/anthropic-ceo-says-it-s-time-to-slow-pace-of-improving-ai-models
[7] Fierce Network, “Exclusive: T-Mobile’s John Saw explains kinetic tokens and why they matter in 6G,” February 23, 2026 — https://www.fierce-network.com/wireless/exclusive-t-mobiles-john-saw-explains-kinetic-tokens-and-why-they-matter-6g;
[8] John Saw, “Why Physical AI Begins with Intelligent Networks,” T-Mobile Newsroom, February 17, 2026 — https://www.t-mobile.com/news/network/why-physical-ai-begins-with-intelligent-networks
[9] CAT NMS Plan, “What are the CAT clock synchronization standards for CAT Reporters?” — https://www.catnmsplan.com/faq/r1; FINRA Rule 6820 (Clock Synchronization) — https://www.finra.org/rules-guidance/rulebooks/finra-rules/6820
[10] Federal Register, July 22, 2025, SR-FINRA-2025-012 (FINRA Rule 6860 timestamp granularity; relief to April 8, 2030) — https://www.federalregister.gov/documents/2025/07/22/2025-13732/; National Law Review summary — https://natlawreview.com/article/finra-aligns-cat-timestamp-requirements-sec-relief-and-extension-2030
[11] CFTC Release 8260-20, “CFTC Orders JPMorgan to Pay Record $920 Million for Spoofing and Manipulation,” September 29, 2020 — https://www.cftc.gov/PressRoom/PressReleases/8260-20
[12] Visa, “Visa and Partners Complete Secure AI Transactions, Setting the Stage for Mainstream Adoption in 2026,” December 18, 2025 — https://usa.visa.com/about-visa/newsroom/press-releases.releaseId.21961.html
[13] Regulation (EU) 2024/1689 (AI Act), Article 73, Reporting of serious incidents (15 days; 2 days for widespread infringement; 10 days for death) — https://artificialintelligenceact.eu/article/73/; Article 12, Record-keeping — https://artificialintelligenceact.eu/article/12/
[14] SDxCentral, “T-Mobile US network chief touts AI transformation and where telecom is best positioned to lead” (Ankur Kapoor, EVP and Chief Network Officer, T-Mobile), August 20, 2026 — https://www.sdxcentral.com/analysis/t-mobile-us-network-chief-touts-ai-transformation-and-where-telecom-is-best-positioned-to-lead/; Kapoor spoke at MFF 2026 on “Silicon, Signal, Service: Earning the Right to Win in the AI Economy” [1]
[15] VentureBeat, “Nvidia’s agentic AI stack is the first major platform to ship with security at launch, but governance gaps remain” (Jensen Huang, GTC keynote, March 17, 2026) — https://venturebeat.com/security/nvidia-gtc-2026-agentic-ai-security-five-vendor-governance-framework
[16] NVIDIA, “Safety for Agentic AI” blueprint (build, deploy and run stages; NeMo Guardrails; continuous monitoring) — https://github.com/NVIDIA-AI-Blueprints/safety-for-agentic-ai; NVIDIA Technical Blog, “NVIDIA-Verified Agent Skills Provide Capability Governance for AI Agents,” May 19, 2026 — https://developer.nvidia.com/blog/nvidia-verified-agent-skills-provide-capability-governance-for-ai-agents/
[17] The Register, “‘Resistance is futile,’ says Qualcomm CEO” (Cristiano Amon, Computex, June 2, 2026) — https://www.theregister.com/personal-tech/2026/06/02/qualcomm-ai-agents-will-be-as-transparent-as-they-will-be-inescapable/5249894
[18] Tom’s Hardware, “Qualcomm roundtable interview transcript” (Kedar Kondap, SVP and GM, Compute and Gaming, Qualcomm), June 4, 2026 — https://www.tomshardware.com/pc-components/cpus/qualcomm-roundtable-interview-transcript-svp-of-compute-and-gaming-talks-snapdragon-c-rtx-spark-and-the-agentic-ai-future; Kondap spoke at MFF 2026 on “Silicon at the Edge of Intelligence” [1]
[19] AI Magazine, interview with Ronnie Vasishta (SVP, Telecom, NVIDIA), May 6, 2025 — https://aimagazine.com/interviews/ronnie-vasishta; Vasishta spoke at MFF 2026 on “Silicon, Signal, Service: Earning the Right to Win in the AI Economy” [1]
[20] SDxCentral, “AT&T sees agentic AI making IoT sexy again” (Shawn Hakl, AT&T Business), March 26, 2026 — https://www.sdxcentral.com/analysis/att-sees-agentic-ai-making-iot-sexy-again/; TelecomTV, “MWC26: AT&T Business on the impact of AI, APIs and telco collaboration,” March 2026 — https://www.telecomtv.com/content/spotlight-on-5g/mwc26-at-t-business-on-the-impact-of-ai-apis-and-telco-collaboration-55092/; title per AT&T newsroom, March 17, 2026 — https://about.att.com/story/2026/cisco-ai-grid-with-nvidia.html; Hakl spoke at MFF 2026 on “Silicon, Signal, Service: Earning the Right to Win in the AI Economy” [1]
[21] SDxCentral, “AT&T CEO John Stankey stays strong on the carrier’s AI-support stance,” July 22, 2026 (Q2 2026 earnings call; line attributed to CFO Pascal Desroches and cited by Stankey) — https://www.sdxcentral.com/news/att-ceo-john-stankey-stays-strong-on-the-carriers-ai-support-stance/; AT&T Q2 2026 earnings transcript — https://investors.att.com/~/media/Files/A/ATT-IR-V2/financial-reports/quarterly-earnings/2026/2Q-2026/t-usq-transcript-2026-07-22.pdf