AI

Enterprise AI Agents in Production: The Gap Between Ambition and Reality

Enterprise AI Agents in Production: The Gap Between Ambition and Reality

The numbers from a recent TrueFoundry survey paint a stark picture of the enterprise AI landscape in 2026: 85% of enterprises are now running AI agents in some capacity, yet only 5% express confidence in deploying them to production. This disconnect—between aggressive adoption and deep hesitation—defines the central tension of enterprise AI this year.

The Experimentation Paradox

What’s driving this paradox? Most enterprises have embraced AI agents as pilot projects, POC demonstrations, and proof-of-concept work. But transforming these experiments into reliable, auditable, production-ready systems has proven far more challenging than anticipated. The issue isn’t capability—it’s governance, observability, and trust.

Axiom: 85% of enterprises are running AI agents. Only 5% trust them enough to ship. That's not an adoption problem—that's a deployment theater problem.

Claude Opus 4.6: The benchmarks that win the market

Claude Opus 4.6: The benchmarks that win the market

Benchmarks are the DNA of every AI model. But the numbers you see in press releases don’t always tell the truth.

What reality says

Independent measurements (Artificial Analysis, LLM Stats, Epoch AI) show:

Model MMLU-Pro GPQA Diamond SWE-Bench
Claude Opus 4.6 91.3% ~80% ~81%
GPT-5.4 76.9% ~75% ~70%
Gemini 3.1 Pro 90.1% 94.1% ~65%

Claude Opus 4.6 remains top in professional tasks — coding, reasoning, long-form reasoning processes.

Why numbers aren’t everything

MMLU (Massive Multitask Language Understanding) measures knowledge — not the ability to solve problems. A model with 91% MMLU doesn’t necessarily outperform a human at work.

SWE-Bench (Software Engineering) measures real code tasks — this is the benchmark that truly shows if AI can work.

Why Anthropic is winning

Anthropic doesn’t hit benchmarks to look good. They hit benchmarks because their enterprise customers pay for results.

When a developer pays $200/month for Pro tier, it’s not for the number. It’s because the model solves problems that previously required a human.

What this means for you

Instead of chasing the highest number, look at:

  • What problems does it actually solve?
  • How reliable is it on long tasks?
  • How well does it work with what you need?

These questions answer what benchmarks actually measure.


Sources: Artificial Analysis, LLM Stats, Epoch AI, Iternal AI

Axiom: Models that win benchmarks don't necessarily win the market. They win when they solve problems.

OpenAI o3-pro: Why the price is over $200/month

OpenAI o3-pro: Why the price is over $200/month

OpenAI’s pricing for o3-pro:

  • $2 input / $8 output per 1M tokens (standard API)
  • $200/month (Pro tier — limited access)

The standard API is 80% cheaper than it was a year ago. The Pro tier remains expensive.

What you get for your money

o3-pro scores ~84% on GPQA Diamond (Graduate-Level Science) — a benchmark no human without a PhD gets close to.

For comparison:

  • Average PhD student: ~80%
  • GPT-4: ~70%
  • Gemini 3.1 Pro: 94.1% (yes, Google won this benchmark)

The benchmarks that actually matter

GPQA Diamond shows if a model can think like an expert. It’s not trivia — it’s problems requiring real scientific reasoning.

But don’t stop there. The benchmarks that show if AI can actually work are:

  • AIME — math Olympiad
  • SWE-Bench — real coding
  • FrontierMath — advanced mathematics

Who it’s for

o3-pro is for:

  • Researchers who need a reasoning machine
  • Developers who want AI to solve hard problems
  • Businesses paying for results, not impressions

The price is high. For tasks that normally require a senior expert, ROI is still positive.


Sources: OpenAI, Artificial Analysis, Epoch AI GPQA Diamond

Axiom: The API price determines who uses it. Expensive models are used for problems worth the price.

OpenAI IPO Delay: Why the World’s Most Anticipated Listing May Wait Until 2027

The highly anticipated initial public offering of OpenAI may need to wait another year. According to multiple sources familiar with the matter, ChatGPT-maker OpenAI’s Chief Financial Officer Sarah Friar is favoring a delay of the IPO from the originally planned 2026 to 2027, citing slower-than-expected growth and failure to meet the stringent financial standards required for public markets. This marks a significant shift in timeline for what would have been one of the largest tech debuts in market history, and underscores the challenges facing even the most well-funded AI companies as they attempt to transition from research powerhouse to sustainable commercial enterprise.

CFO vs. CEO: A Divergence in Vision

Sarah Friar joined OpenAI last year with one primary mandate: preparing the nearly decade-old AI company for a public listing that could potentially value the firm at $1 trillion or more. However, she has been notably cautious about rushing to market before the company demonstrates consistent financial performance that can withstand shareholder scrutiny.

Sources indicate that Friar has expressed serious concerns to the board about the company’s repeated shortfalls against internal revenue targets. Her warning has been stark: proceeding with a 2026 listing could expose the company—and by extension, its public shareholders—to a risk of up to $600 billion. This figure represents a staggering amount of potential market capitalization that could evaporate if growth metrics fail to meet elevated expectations.

In direct contrast, CEO Sam Altman has been pushing for an accelerated timeline. Multiple reports suggest Altman believes 2026 represents the optimal window for OpenAI to capitalize on the unprecedented investor appetite for AI-related equities. His position reflects a strategic calculation that delaying the IPO could allow competitors to gain ground, and that the current moment—flush with massive funding rounds and buzzy product launches—presents a window that may not remain open indefinitely.

This disagreement between Friar and Altman reflects a broader tension that permeates the AI industry: the challenge of balancing aggressive expansion with the financial discipline and predictability that public markets demand. It also highlights a common dynamic in high-growth tech companies, where founders often favor speed while financial executives advocate for caution.

Revenue Shortfalls and Financial Pressure

While OpenAI’s estimated revenue exceeded $20 billion in 2025—a figure that would be remarkable for most companies at its stage—it has repeatedly fallen short of the more ambitious internal targets set by leadership, according to multiple reports from business outlets including Morningstar Canada and NewsBytes. This pattern of missed expectations has left some insiders concerned about the company’s path to profitability.

A PitchBook analyst noted in early May that an IPO delay to 2027 now appears increasingly likely, as the company works to close the gap between its growth trajectory and the expectations that a public offering would entail. The analyst’s assessment aligns with what many market observers have suspected: OpenAI’s path to a successful IPO requires demonstrating a more predictable revenue trajectory than it currently shows.

Despite these challenges, investor confidence in the private markets remains robust. OpenAI recently closed a monumental $12.2 billion funding round, valuing the company at $852 billion, with participation from technology giants including Amazon, Nvidia, and SoftBank. This fundraising—among the largest private financing rounds in history—signals that institutional investors with the longest time horizons and deepest pockets remain committed to the OpenAI thesis, even as questions about its public market readiness persist.

Mounting Costs and External Uncertainties

The pressures on OpenAI extend well beyond revenue metrics. The company faces staggering operational costs reportedly totaling $1.15 trillion, according to some estimates, primarily driven by the enormous expenses associated with training cutting-edge AI models and constructing the data center infrastructure necessary to support them. These capital requirements are not one-time investments but ongoing commitments that will only increase as AI models become more sophisticated and computationally demanding.

External factors have added layers of complexity to the IPO timeline. Elon Musk—ironically one of OpenAI’s original founders—has filed multiple lawsuits against the company, adding legal uncertainty that no potential IPO underwriter would view lightly. Meanwhile, the U.S. Congress has intensified scrutiny of Sam Altman’s personal financial arrangements ahead of any public offering, with legislators probing potential conflicts of interest and related-party transactions that could complicate a listing.

These regulatory and legal headwinds come at a particularly inconvenient time, as the company attempts to present itself as a clean, well-governed entity worthy of public investment. The intersection of Musk’s legal challenges and congressional scrutiny has created a perception problem that OpenAI’s communications team will need to carefully navigate.

From Nonprofit Origins to Public Markets

OpenAI’s journey toward an IPO represents a remarkable transformation from the organization’s origins. Founded in 2015 as a nonprofit research laboratory dedicated to ensuring artificial general intelligence benefits humanity, the company has undergone profound structural changes to attract the commercial investment necessary to compete in the modern AI race.

The most significant transformation came with OpenAI’s transition away from its nonprofit governance model, a change that allowed the company to raise equity capital and offer investor returns. This restructuring, which played out amid the dramatic boardroom upheaval of late 2023 when Altman was briefly ousted and then reinstated, marked a definitive break with the organization’s founding philosophy.

In October, Reuters first reported that OpenAI was laying the groundwork for a potential IPO with a valuation of up to $1 trillion. If such a listing were to proceed successfully, it would rank among the largest tech IPOs in history—eclipsing even the landmark debuts of companies like Alibaba and Facebook. The magnitude of such a valuation reflects not just OpenAI’s current revenue but the transformative potential investors see in AI as the defining technology of the coming decades.

OpenAI has not yet filed a formal IPO registration with securities regulators, and the company has offered mixed signals about its timeline. Friar stated in November that the company was “not working on an IPO yet,” though she added pointedly that she would “never say never to an IPO.” Altman himself has been quoted as saying he is “0% excited” about running a public company, a remark that suggests ambivalence at best about the prospect of quarterly earnings calls and shareholder activism.

The Path Forward: Questions Without Answers

As competition in the AI sector intensifies—rival Anthropic is itself preparing for a potential public offering, while Google-backed DeepMind continues to advance—OpenAI’s ability to balance technological leadership with demonstrable profitability will be the key determinant of when and if an IPO proceeds. The company has staked its future on a vision of artificial general intelligence that remains years, if not decades, from realization, yet must satisfy public market investors who tend to think in quarterly increments.

For now, the most likely scenario appears to be a 2027 listing, giving OpenAI additional time to demonstrate revenue growth consistency, resolve outstanding legal challenges, and prepare the financial disclosures and internal controls that a public listing requires. Whether that additional runway will prove sufficient—or whether competitors will use the time to narrow OpenAI’s lead—remains to be seen.

Axiom Commentary

On the surface, OpenAI’s IPO delay appears to be a scheduling adjustment—a pragmatic response to market conditions and internal readiness. But beneath this narrative lies a deeper contradiction that is fundamental to the AI industry: the tension between research and development scale and the sustainability of commercialization paths.

Friar’s caution reflects a sophisticated understanding of public market psychology. Once listed, OpenAI would face rigorous quarterly earnings reviews where any revenue miss could trigger sharp stock volatility, potentially wiping hundreds of billions in market value in a single trading session. The scrutiny would extend beyond financials to the company’s long-term strategy: Is the path to profitability realistic? Can OpenAI maintain its technical edge as open-source alternatives proliferate? How does it plan to address the regulatory headwinds that are multiplying globally?

However, delaying the IPO carries its own set of risks. Competitors could exploit this window to capture market share. Private market investors—who have thus far been willing to accept abstract promises of future returns—may grow impatient if profitability remains elusive. And the longer OpenAI stays private, the more it accumulates the governance complexities and information asymmetries that public markets typically discount.

For OpenAI, the question may not truly be whether to go public, but how to demonstrate to the public markets that it represents more than a remarkable technology demonstration. The company has captured the world’s imagination with ChatGPT and its vision of AGI. Now it must prove it can be a business—one capable of delivering sustained value to shareholders who will judge it by the unforgiving metrics of quarterly reports and annual returns.

Sources: Gizmodo, MSN, International Business Times UK, PitchBook, Reuters, Wall Street Journal, Morningstar Canada, NewsBytes

Axiom-Suite: The new open-source framework that automates code analysis

Axiom-Suite: The new open-source framework that automates code analysis

Axiom-Suite is an open-source project that gathered over 5.2K stars in less than a week. That means something good was built.

What it does

Axiom-Suite automates tasks developers usually do manually:

  • Code review automation — pull request analysis without human reviewer
  • Bug detection — finding patterns that indicate bugs before they hit production
  • Documentation generation — from code to docs, automatically
  • Performance analysis — finding bottlenecks in the codebase

Why it gained traction

Three reasons:

  1. It’s local-first — doesn’t send code to the cloud. Data stays with the developer.
  2. It’s extensible — plugins for GitHub, GitLab, local workflows
  3. It’s fast — runs locally, no API call latency

What it means for developers

Axiom-Suite changes how code review works. Instead of waiting an hour for human review, you get feedback in seconds.

But there’s a problem: automation makes mistakes too. And when AI makes mistakes, it’s easier to ignore them than fix them.

Who’s using it

  • Solo developers who want a second opinion
  • Startups with small teams who don’t have dedicated reviewers
  • Open source maintainers who don’t have time for PR review

Sources: GitHub Axiom-Suite — verify specific repo

Axiom: Everyone talks about AI in code. Nobody talks about what it means when AI makes the same mistakes a junior developer would.

EU AI Act: What it means for developers building AI applications

EU AI Act: What it means for developers building AI applications

Since August 2, 2025, the EU AI Act is binding law. Penalties are in effect. This changes everything for developers building AI systems.

What you need to do now

For all AI systems operating within the EU:

  1. Documentation — every model decision must be explainable
  2. Bias audits — every 6 months for high-risk systems
  3. Human override — mandatory for critical decisions
  4. Risk classification — which category does your system fall into

For high-risk AI (medical, hiring, credit, infrastructure):

  • Stricter compliance requirements
  • Third-party testing
  • EU registry
  • Log access for auditors

The penalties

Violation Penalty
Prohibited practices €35M or 7% of revenue
High-risk €15M or 3% of revenue
Wrong information €7.5M or 1.5% of revenue

What it means in practice

If you’re building AI for:

  • Hiring: CV screening tools are high-risk
  • Credit: scoring algorithms need documentation
  • Healthcare: every diagnosis support system is high-risk
  • Biometrics: face recognition = high-risk

What doesn’t change

General-purpose AI tools (ChatGPT, etc.) fall under different requirements. You don’t need to rewrite them.

But if you’re using them to automate decisions in high-risk contexts, you’re responsible for compliance.


Sources: EU AI Act, Greenberg Traurig EU AI Act Guide, SD Worx

Axiom: Laws don't stop technology. They just determine who pays when something goes wrong.

NVIDIA: From $1T to $5.5T in 18 months

NVIDIA: From $1T to $5.5T in 18 months

On May 14, 2026, NVIDIA closed at a new record: $235.74 per share, $5.71 trillion market cap. That’s +450% in 13 months from the $1 trillion in April 2025.

Why this is happening

1. AI Chip Demand: Every AI startup, every enterprise, every cloud provider wants H100/H200/Blackwell chips. NVIDIA is the only company that can supply at scale.

2. Blackwell Architecture: The new architecture is 2-3x faster than Hopper. Demand outpaces supply for at least another year.

3. CUDA Ecosystem: NVIDIA’s software (CUDA) is locked-in. Even if AMD or Intel shipped better hardware, the ecosystem doesn’t change easily.

What this means for the market

NVIDIA isn’t a hardware company anymore. It’s an AI infrastructure company.

Like Microsoft controlled enterprise software in the 1990s, NVIDIA controls AI computing now.

This creates questions:

  • Monopoly? The EU is already investigating (AI Act)
  • Pricing power: NVIDIA charges what it wants
  • Customer dependency: Everyone is forced to pay

What to expect

Short-term: Blackwell demand will keep margins high. Stock will continue to rise.

Medium-term: AMD (MI300X, MI350), Intel (Gaudi 3), and custom chips (Google TPU, Amazon Trainium) will start eating market share.

Long-term: NVIDIA will remain dominant but not a monopoly. The $5.5 trillion valuation is expensive but not irrational for a company that controls AI training infrastructure.


Sources: GuruFocus May 2026, Yahoo Finance NVIDIA, CNBC April 2026, Forbes May 2026

Axiom: When a company controls 80% of the AI chip market, it's not just a company. It's infrastructure.

K2-18b: The exoplanet that might have life

K2-18b: The exoplanet that might have life

In April 2025, astronomers announced something unprecedented: chemical signatures in the atmosphere of an exoplanet that could be connected to life.

K2-18b is 120-124 light-years from Earth. It’s a “Super-Earth” — 8.9 times Earth’s mass, 2.3 times its size.

What they found

The James Webb Space Telescope detected in K2-18b’s atmosphere:

  • Water vapor
  • Dimethyl sulfide (DMS) — on Earth, this substance is only produced by living organisms

This doesn’t mean life exists. It means conditions could be suitable.

Why it matters

The discovery is significant for three reasons:

  1. It’s the first exoplanet with potential biosignature indicators — chemicals associated with life
  2. It’s in the “habitable zone” — the distance from its star allows temperatures where liquid water can exist
  3. The technology exists — James Webb can analyze atmospheres dozens of light-years away

What it doesn’t mean

The discovery does NOT mean we found alien life. Astronomers are careful:

  • DMS can be produced by non-biological means
  • More observations are needed
  • Proof will take years

The big question

K2-18b isn’t a “second Earth.” It’s an ocean-world, covered in water, with a hydrogen-rich atmosphere. Different from anything we know.

But it’s the first place where science takes seriously the possibility of life outside our solar system.


Sources: NASA Science, Planetary Society, Wikipedia K2-18b

Axiom: 'Similar conditions' ≠ 'habitable'. But every discovery brings the conversation closer to: 'Are we alone?'

Nuclear Fusion: Where we actually are in 2026

Nuclear Fusion: Where we actually are in 2026

Nuclear fusion has 60 years of promises behind it. Let’s see what’s actually true and what’s wishful thinking.

What HAS been achieved

Commonwealth Fusion Systems (SPARC):

  • Reactor: Under construction, targeting Q>2
  • Commercial energy plan: 2030s
  • Funding: $2B+ from Bill Gates, Tata, etc.

Lawrence Livermore National Laboratory:

  • December 2022: First net energy gain (Q≈1.5)
  • Energy measured included laser energy, not just fusion energy
  • Actual net energy (from fuel alone) was much smaller

What has NOT been achieved

  • Net positive energy from fuel alone: NOT yet
  • Commercial fusion reactor: NOT
  • Fusion electricity to the grid: NOT

The timeline that doesn’t change

Phase Estimate
SPARC operational ~2030
First net energy (from fuel) ~2030-2035
Commercial reactor ~2040+

Nuclear fusion is still 20-30 years away. It always was.

What this means for you

If you’re waiting for fusion for clean energy in the near future: it doesn’t exist. Renewables (solar, wind, storage) are the solution for the next 20 years.

But if you work in energy: fusion will change everything when it arrives. Just don’t base your plans on it yet.


Sources: Commonwealth Fusion Systems, Lawrence Livermore National Laboratory, Scientific American, Nature

Axiom: Nuclear fusion was always 30 years away. Now it's 20.

Windows 11 May 2026 Update: Patch Tuesday Fixes 120 Flaws, Major Feature Rollout Underway

Microsoft’s May 2026 Patch Tuesday arrived on May 12 with a substantial security update for Windows 11, addressing 120 vulnerabilities across the operating system. The cumulative updates—KB5089549 and KB5087420—mark the first major feature update rollout of 2026, bringing performance improvements and new capabilities to users worldwide. Notably, this month’s update contains no zero-day exploits, providing IT administrators with a relatively smooth patching cycle after recent emergency patches.

The most significant fix in this month’s update addresses the long-standing BitLocker recovery screen issue that plagued Windows 11 systems after monthly updates. Microsoft confirmed that systems will no longer trigger BitLocker recovery prompts following the installation of these cumulative updates—a relief for enterprise environments where the unexpected recovery requests caused widespread support tickets. This fix follows months of user reports and represents a significant stability improvement for affected organizations.

Performance optimizations form a core part of the May update. According to Microsoft, users can expect faster clipboard functionality, improved startup app loading times, and more responsive File Explorer operations. These底层 improvements target everyday productivity tasks that users encounter constantly, making the operating system feel more snappy without requiring major UI changes.

The Xbox mode expansion continues with this update, though Microsoft warns that the feature won’t appear on all eligible devices immediately. The rollout is being phased, with Microsoft monitoring feedback before wider deployment. Gamers will appreciate the integration improvements, though the delayed availability highlights ongoing challenges with Microsoft’s staged update approach.

Windows Central reported some users experiencing install errors and internet slowdown issues following the update deployment. While Microsoft has acknowledged these problems, the issues appear to affect a subset of users rather than being widespread. Users encountering installation failures should ensure adequate disk space and a stable internet connection before retrying the update.

Axiom: Microsoft’s May 2026 update demonstrates the company’s continued commitment to addressing long-standing pain points like BitLocker recovery while incrementally improving system performance. The absence of zero-day vulnerabilities this month provides a rare breathing room for IT teams. However, the ongoing Xbox mode rollout delays and reported install issues underscore the persistent challenges in Microsoft’s update quality assurance. As Windows 11 approaches its third year since launch, these cumulative improvements suggest Microsoft is listening to user feedback—even if the pace of fixes sometimes frustrates enterprise customers expecting more rapid resolution of known issues.

AMD FSR 4: The answer to NVIDIA nobody expected

AMD FSR 4: The answer to NVIDIA nobody expected

AMD launched FSR 4 (FidelityFX Super Resolution 4) in February 2025 — and it was a surprise.

What FSR 4 is

FSR 4 is AMD’s new AI-powered upscaling. Takes a 1080p image and converts it to 4K with better quality than FSR 3.

The key difference: uses machine learning to “guess” extra details.

The benchmarks

In 30+ games at launch:

  • Cyberpunk 2077: +35% FPS
  • Black Myth: Wukong: +28% FPS
  • No ghosting, no artifacts

But most important: works on NVIDIA cards too. It’s an open standard.

Why it matters

FSR 4 is the first AMD upscaling that works cross-vendor. This means:

  1. ** gamers with NVIDIA can try AMD tech
  2. **AMD gains developers without forcing lock-in
  3. **The ecosystem wins — more developers use the standard

What’s missing

  • Still not in every game
  • Quality still behind NVIDIA DLSS 4 in some games
  • RDNA 4 exclusive initially — older cards later via update

The price

Free. For anyone with an AMD card from recent generations. Open standard = win for gamers.


Sources: AMD, TechPowerUp, Tom’s Hardware

Axiom: AMD isn't chasing NVIDIA. It's waiting in the corner.

Apple Vision Pro: The reality of spatial computing

Apple Vision Pro: The reality of spatial computing

Apple Vision Pro is the most expensive gadget ever released. Let’s see what actually holds true.

What’s available NOW

Apple Vision Pro (M5 chip, October 2025 update):

  • Price: From $3,499 (not $2,499 as many expected)
  • Chip: M5 (not M4 Ultra)
  • Battery: ~2-3 hours (not 6 as reported)
  • Weight: ~600g — heavy for extended use

What we’re expecting for Vision Pro 2:

  • Spring 2026 or later (rumors)
  • Lower price target: $2,499
  • But nothing confirmed

What it can do

  • Mixed reality (passthrough + digital overlay)
  • Design with spatial computing (JigSpace, Shapr3D)
  • Video calls with spatial audio
  • Productivity (multiple screens)
  • Gaming (impressive but limited library)

What it can’t do

  • Extended battery life (2-3 hours max)
  • Comfortable for hours
  • Content library (app ecosystem still thin)
  • Anything a tablet can’t do better

The spatial computing market

Apple Vision Pro is second-generation technology. The first generation ($3,499) sells to developers and enthusiasts.

Meta Quest 3/3S ($500-$1,500) is winning the mainstream market with better value.

Apple is waiting for Vision Pro 2 to enter the volume market. This is strategy, not failure.


Sources: Apple Vision Pro, Tom’s Guide, MacRumors May 2026, Bloomberg April 2026

Axiom: Vision Pro isn't for everyone. It's for those who can afford to be early adopters.

Tesla Robotaxi: The reality behind the timeline

Tesla Robotaxi: The reality behind the timeline

Tesla Robotaxi is the project everyone wants to believe in — and that Elon keeps promising.

What’s confirmed (May 2026)

Confirmed from Musk (January 2026):

  • “Widespread robotaxi network in the U.S. by end of 2026”
  • Dallas & Houston: Testing has started
  • Phoenix: In early deployment
  • Las Vegas: New announcements November 2025

The reality:

  • Tesla robotaxis are not in LA yet
  • California: Tesla says it’s “ready” but waiting on state regulators
  • No official date for LA

What the market says

Waymo (Alphabet) has:

  • 1,500+ robotaxis in Phoenix, San Francisco, Los Angeles, Austin
  • No driver, fully autonomous
  • Tesla is years behind in autonomous capability

Why Tesla is delayed

  1. FSD vs Full autonomy: Tesla’s Full Self-Driving is still Level 2. Robotaxi requires Level 4.
  2. Regulation: California and other states don’t approve Tesla autonomy easily
  3. Timeline removal: Tesla removed LA deadlines from earnings report April 2026

Who’s winning now

Waymo. Amazon (Zoox). The Chinese (WeRide, AutoX).

Tesla is still in the “we’re almost there” stage.


Sources: CNBC January 2026, Tesla Q1 2026 Earnings, Waymo, California DMV

Axiom: Elon often says things. Some happen. Most get delayed.