Purplelink Daily Digest #7 — June 28, 2026

DeepSeek's DSpark applies speculative decoding to their production inference stack, with the paper reporting concrete throughput and latency numbers for large MoE models — a configuration where speculative decoding gains are harder to achieve than in dense models. The non-obvious angle: MoE architectures have irregular compute patterns that typically undermine draft-model alignment, so any reproducible speedup here has direct implications for cost-per-token at scale. Researchers running vLLM or SGLang on DeepSeek-class models should benchmark against the reported figures.

Anthropic's export restrictions on its Mythos-class models are creating a direct market opening for Asian frontier labs, with multiple startups shipping capability-comparable alternatives into markets Anthropic cannot serve. The strategic implication is that export controls intended to preserve US AI advantage may be accelerating capability diffusion by forcing customers to adopt non-US alternatives. This is a concrete data point for anyone modeling the long-term competitive dynamics of AI export policy.

A practical RDMA cluster configuration guide for AMD Strix Halo (the APU with 128GB unified memory) running vLLM fills a real gap: most inference cluster documentation assumes discrete GPU nodes, not high-memory APU configurations. For a one-person macOS/iOS shop or small lab wanting to run 70B-class models without discrete GPU infrastructure, Strix Halo's unified memory bandwidth makes it a credible alternative — but RDMA multi-node setup has been underdocumented until now. The guide's reproducibility depends on specific firmware and ROCm versions, which should be verified before production deployment.

The GRU-linked campaign has evolved beyond account takeover: attackers now specifically harvest Signal Backup Recovery Keys, enabling full historical message restoration on attacker-controlled devices. The non-obvious threat is that a single key exfiltration is permanent — unlike a session hijack, it grants retroactive access to the entire encrypted archive. Defenders building secure comms policies for high-value targets should treat the recovery key as a credential requiring the same protection as a private key.

The SSU/FBI joint disclosure details a sustained smishing campaign impersonating telecom support to harvest OTPs and linking codes from Ukrainian government officials, military, and activists. The campaign's longevity — described as 'long-running' — suggests operational security failures at the target population level, not just technical ones. Connects to: FBI Warns Russian Intelligence Hackers Target Signal Backup Recovery Keys.

A malicious payload hidden in a repository remains invisible to static scanners, AI agents, and human reviewers until the agent executes setup steps — at which point the payload runs with the agent's ambient permissions. This is a supply-chain attack vector that scales with agentic coding adoption: the attack surface grows every time a developer delegates 'clone and set up this repo' to an autonomous tool. The specific invisibility mechanism is not disclosed in the report, making independent reproduction and defensive tooling difficult.

Kaspersky's StrikeShark tracking identifies SharkLoader as a previously undocumented loader purpose-built for Cobalt Strike Beacon delivery, suggesting a threat actor investing in custom tooling rather than commodity loaders. Custom loaders typically indicate a more disciplined operator trying to evade signature-based detection of known loader families like BruteRatel or Donut. The absence of attribution details in the disclosure limits threat-intel correlation.

Google is capacity-constraining Meta's Gemini API access, not for policy reasons but because it cannot provision enough compute — a rare public signal that frontier inference capacity is genuinely supply-constrained at the hyperscaler level. The implication for AI infrastructure economics is that even well-capitalized customers cannot simply buy their way to unlimited frontier model access, which creates durable pricing power for whoever controls TPU/GPU supply. This also explains why Meta continues to invest heavily in its own inference infrastructure rather than relying on third-party APIs.

Navan is up 30% YTD while IGV is down 15% and software forward multiples have dropped below the S&P 500 for the first time — the post argues Navan's fintech-adjacent revenue model (transaction fees on travel spend) is structurally more defensible than pure SaaS subscription revenue in an AI-disruption environment. The non-obvious point is that software businesses with embedded financial flows are being repriced upward relative to pure workflow software, which has direct implications for how indie and startup founders should think about monetization architecture. The analysis is qualitative and lacks a rigorous multiple comparison, so the causal claim deserves scrutiny.