The mythos of Mythos and Allbirds takes flight to the neocloud

Summary of The mythos of Mythos and Allbirds takes flight to the neocloud

by Practical AI LLC

45m•April 23, 2026
Listen to Episode

Overview of The mythos of Mythos and Allbirds takes flight to the neocloud

This episode of the Practical AI Podcast (hosts Daniel Whitenack and Chris Benson) discusses recent AI-business headlines and how they reflect shifting markets, technical trends, and governance challenges. Major stories: Allbirds’ surprising pivot from footwear to AI compute ("neocloud"), the concept of NeoCloud / AI-native cloud infrastructure, Anthropic’s rumored frontier model “Mythos” and Project Glasswing (security testing), the emergence of “token maxing” as a developer/engineering behavior, and a recent federal ruling that AI chat logs can be discoverable in litigation. The hosts use these examples to draw practical lessons about risk, governance, economics, and operational strategy for teams using or building AI.

Key topics discussed

Allbirds pivot → NeoCloud

  • Allbirds sold its footwear assets (to American Exchange Group) and repositioned the remaining corporate shell into AI compute infrastructure by buying GPUs.
  • Market reaction: shares jumped dramatically (the hosts cite ~700%).
  • Context: this is symptomatic of companies pivoting into AI compute; raises questions about domain expertise and whether modest capital (~$50M mentioned) is meaningful in the large AI data-center market.

What is a “NeoCloud” (AI-native cloud)?

  • NeoCloud (aka AI-native cloud): cloud infrastructure purpose-built for AI workloads (GPU-first, tuned for training/inference, data movement patterns).
  • Examples: CoreWeave, Together AI, Lambda Labs.
  • Differences from hyperscalers: specialization vs broad general-purpose services; potential advantages for AI-first teams but also competition and scale dynamics with hyperscalers (AWS/GCP/Azure).

GPU & chip supply constraints

  • GPU / AI chip supply is concentrated (NVIDIA, TSMC, AMD, Intel, Apple, Qualcomm called out). If many companies pivot to data centers, chip supply could become more strained.

Edge/embedded vs centralized compute

  • Dual trend: expansion of centralized GPU data centers AND growth of embedded/far-edge AI (on-device, kiosks, manufacturing floor, cars).
  • Hosts expect both to grow; long-term strategic differentiation may come from embedded/edge specialization.

Anthropic’s Mythos and Project Glasswing

  • Mythos: rumored next-generation Anthropic frontier model (successor to Opus/Claude Code).
  • Reported capability: very effective at finding/exploiting security vulnerabilities across software stacks.
  • Project Glasswing: closed security-review project where Anthropic allowed select companies (reportedly ~40, ~12 public) to test systems against Mythos so vulnerabilities could be fixed before public exposure.
  • Implication: advanced frontier models could dramatically raise both defensive and offensive cybersecurity stakes. Anthropic’s handling also signals a marketing/governance angle.

“Token maxing” and developer behavior

  • “Token maxing”: intensive consumption of LLM/API tokens to accelerate developer productivity; has been gamified in some firms (scoreboards, leaderboards).
  • Meta example: public leaderboard and token-gaming behavior.
  • Hosts note tension: token usage can be useful but is sometimes a vanity metric; organizations must learn effective metrics and guard against waste.

Legal discovery and AI chat logs

  • A federal judge ordered production of a defendant’s AI chat outputs (Claude) used in legal prep — reinforcing that AI chats are likely treated as third-party communications (not attorney-client privileged).
  • Practical consequence: chat logs with third-party AI services may be discoverable; firms should avoid putting sensitive/confidential material into third-party models without safeguards.

Main takeaways

  • NeoClouds are real and gaining traction, but building meaningful AI compute capacity is capital- and expertise-intensive; pivots (e.g., Allbirds) may be noisy marketing plays more than durable advantages.
  • Frontier models like Mythos could shift the cybersecurity landscape (both offensive and defensive). Responsible gating and governance will be increasingly important.
  • GPU supply constraints remain a structural limiter; if many players chase compute, scarcity and price/availability pressures will intensify.
  • Token consumption is a new operational lever for productivity but can be inefficient, gamed, or a vanity metric; teams need to define meaningful productivity/ROI metrics.
  • Legal and compliance risk: interactions with third-party AI providers are potentially discoverable and not protected as privileged communications; contracts and processes should reflect that reality.
  • Edge/embedded AI is a major growth opportunity that may offer differentiation versus competing in centralized compute.

Notable quotes & paraphrases

  • “What do you do with a bunch of cash but buy GPUs?” — highlighting the fever around buying compute.
  • “NeoCloud = cloud infrastructure built specifically for AI workloads (GPU-first).”
  • “Mythos was reportedly particularly adept at uncovering security vulnerabilities… Anthropic started Project Glasswing to give select companies time to fix those issues.”
  • “Token maxing is a vanity metric — correlation isn’t causation: spending tokens doesn’t automatically mean you’re productive.”
  • “AI systems are not lawyers…conversations with AI are treated like a third party,” meaning chats can be discoverable.

Recommended actions (for founders, engineering leaders, legal/compliance teams)

  • For engineering leaders:

    • Track token usage ROI: measure outputs tied to business outcomes (not just tokens consumed).
    • Avoid letting token scoreboards become morale/vanity traps; encourage efficient prompting and tooling.
    • Assess whether to use specialized NeoCloud providers vs hyperscalers depending on workload, latency, cost and vendor maturity.

  • For executives / strategy:

    • Be skeptical of headline pivots into compute — evaluate domain expertise, capital needs, and chip supply constraints before assuming success.
    • Consider hybrid strategies: leverage centralized GPU capacity for large models and invest in edge/embedded where product differentiation exists.

  • For security teams:

    • Assume advanced models can both find and create exploits; treat model-assisted security testing as a double-edged sword.
    • Plan for adversarial scenarios and invest in AI-aware security monitoring and controls.

  • For legal / compliance:

    • Update policies and contracts: explicitly warn employees/clients not to put privileged/confidential material into third-party AI services.
    • Treat AI chat logs as potentially discoverable; preserve logs and consider private, auditable deployments for sensitive work.
    • Consult counsel on jurisdictional differences and incorporate explicit clauses in vendor/customer agreements.

  • For product teams:

    • If privacy/confidentiality is core, evaluate private model deployments (on-prem or VPC-hosted), no-log options, or bespoke secure APIs.
    • Track regulatory and court developments closely.

Quick glossary / items to follow

  • NeoCloud / AI-native cloud: cloud services tuned for large AI workloads (GPU-first).
  • Mythos: rumored Anthropic frontier model (post-Opus).
  • Project Glasswing: Anthropic’s closed security review program for Mythos.
  • Token maxing: maximization of LLM API token consumption by developers/teams.
  • Opus / Claude / Claude Code: Anthropic’s prior models and coding assistant integrations.
  • CoreWeave, Together AI, Lambda Labs: companies often cited as NeoCloud or AI infrastructure providers.

If you want to dive deeper into any of these topics (e.g., how to build internal guardrails for token usage, or a checklist for moving confidential workflows off third-party LLMs), the episode and hosts offer a practical orientation from which to plan next steps.