OM DENNE EPISODE
Today's theme is the paper trail. Microsoft's outage now has a price tag, though still no root cause. The Nvidia story has moved further from the personal scandal toward evidence: who financed which GPU purchases, and which chips trained which Chinese models. The one genuinely new security story is a Russian hosting provider hit by a DDoS attack while it is still rebuilding from a destructive breach.
---
Infrastructure: Microsoft outage gets a cost estimateFollow-up. The outage, its scope and Microsoft's two explanations were covered Nov 14–15. This item covers what's new.
- First loss estimate. A Morningstar analyst puts insured losses at $300–500M, preliminary. For comparison, the 2024 CrowdStrike incident cost up to $1.5B.
- That is a lower bound on the real cost. Much of the disruption, such as trading desks reverting to phone and fax, won't appear as insured losses.
- Track record. Ookla counts this as Microsoft's second major outage of 2026 and its longest so far. It follows a multi-day authentication outage in September.
- Still open. There is no root cause yet. Microsoft has said only "servers unexpectedly unavailable" and problems during a network capacity expansion. It has promised a full incident report after internal reviews. Until then, its denial of any malicious or AI-agent involvement is Microsoft's own account.
What to watch: Whether the post-incident report names a specific change-management failure. Two long outages in one quarter will bring scrutiny to Microsoft's deployment safeguards, not only to this event.
---
Nvidia: the after-Huang pictureFollow-up. The resignation, Kress as interim CEO, Harvey Jones as chair, the market damage and the Megaspeed allegations were covered Nov 10–15. New items only.
Political and regulatory threads- Congressional inquiry. Rep. Richard Neal and Sen. Ron Wyden opened an inquiry into how the administration handled Nvidia's China sales. They note that H200 sales to China were approved in December 2025, reportedly after officials had been briefed on the Huang–Alice Huang relationship in late 2025.
- Senate Banking. Sen. Tim Scott has asked Huang and Commerce Secretary Howard Lutnick to testify. Reports suggest Lutnick's job could be at risk over Commerce's export oversight.
- Huang's testimony. He has agreed to testify before Congress before Thanksgiving, likely in closed session.
- Scale of the alleged channel. Megaspeed reportedly bought about $4.8B in Nvidia chips over two years. A related firm, Shanghai Century Litong, reportedly became a preferred cloud partner and facilitated about $7.5B in purchases, including banned chips. Megaspeed denies wrongdoing.
- China: Calls the spy allegations a groundless smear. An estimated $3B+ in Chinese AI compute projects that depend on restricted Nvidia chips now face uncertainty.
- Singapore: Had seized Alice Huang's passport in 2025. It found insufficient evidence for false-declaration charges, but warned her and barred her from re-entering. The probe into Megaspeed and six others continues, and Singapore says US cooperation has been limited.
- South Korea: Nvidia reaffirmed its 260,000-GPU supply commitment. The government's AI strategy council held an emergency meeting anyway.
- Israel: Nvidia's largest R&D hub outside the US, about 6,000 staff and built on Mellanox. Local VCs see an opening for inference-chip challengers but say leadership turmoil alone won't move customers.
- Taiwan and Japan: The Taipei HQ remains on track. Japan says it is monitoring the situation.
- Reinstatement petition. It claims 115k+ signatures, including Elon Musk, Garry Tan, Tim Sweeney and Amjad Masad. It also claims about 6,000 employees and 63,000 shareholders signed. All of these figures are self-reported and unverified.
- Permanent CEO. Names floated include Apple's Johny Srouji and Nvidia's Michael Kagan. The list is explicitly tentative.
- Security leadership. CSO David Reber stepped down in August for health reasons and was replaced by Saša Zdjelar. Nothing public links this to the Megaspeed matter.
Next date: Q3 FY2027 earnings are tomorrow, Tue Nov 17. This is Kress's first call as interim CEO.
---
Were Chinese frontier models trained on banned Blackwells?Former OpenAI researcher Will DePue analyzed open-weight models from DeepSeek, Moonshot AI, MiniMax and Ant Group. He claims they show "fingerprints" of training on export-banned Blackwell GPUs, naming DeepSeek V4 and Moonshot's Kimi K3 specifically.
Caveats:
- The analysis has not been independently verified. The sources don't describe the method in detail, and inferring training hardware from released weights is an unusual claim that needs scrutiny.
- Even if the claim is correct, it doesn't prove an export violation. Training in offshore data centers such as Malaysia is not necessarily illegal under current rules.
Context:
- Epoch AI estimates about 2M advanced chips entered China through concealed routes in 2024–25.
- Epoch AI also estimates about 275k Nvidia accelerators went to Malaysia over the same period.
- About 600k H200s are reportedly stockpiled in Hong Kong awaiting Chinese import clearance. Those chips are US-approved for sale. Why they are held up on the Chinese side is unclear.
Why it matters: If the fingerprinting holds up, export controls would have failed to stop frontier training, not merely slowed it. That would arrive at the same time as congressional pressure over the Megaspeed channel.
---
The "circular financing" allegationA major financial newspaper, relying on anonymous sources, reports the following:
- The core claim. Nvidia invested in AI startups and smaller clouds that then bought its GPUs. The alleged goal was to sustain the appearance of demand and absorb oversupply ordered from TSMC.
- The August meeting. CFO Colette Kress argued for scaling back these arrangements and was harshly rebuffed by Huang.
- Business pressures cited:
- Hyperscalers are shifting to custom silicon.
- Some Nvidia-funded smaller clouds reportedly couldn't afford the GPUs they had committed to.
Where accounts differ:
- The OpenAI number. This report describes a "$250B deal with OpenAI" that Kress wanted to scale back. Other reporting says $250B was a discussed credit commitment for OpenAI's Ohio campus, which Kress successfully pushed down to $105B. The $105B figure matches the guarantee we reported Nov 14. "$250B deal" overstates something that apparently never closed at that size.
- The tone of the clash. Separate anonymous-source reporting has Huang calling Kress "a deer in the headlights" this summer. It says she also questioned credit-guarantee deals for data-center operators involving Huang's son-in-law.
- What the reports share. Both point to a CFO who resisted the financing strategy and lost the argument at the time.
- What it adds up to. Read together, the reports suggest a substantive strategy dispute was already underway before the scandal. Nothing yet confirms the stronger claim that demand was deliberately manufactured.
Online speculation. Claims about RICO or antitrust exposure are just that. No legal action has been announced.
Startup funding impact.
- Nvidia joined 105 funding rounds this year, according to PitchBook.
- Its deals include an $800M check into Reflection AI, stakes in Mistral, Crusoe, Lambda and Black Forest Labs, and a pending $13B Hugging Face acquisition expected to close in 2027.
- VCs expect Kress to write smaller, fewer checks, with neoclouds and pre-revenue infrastructure most exposed.
- Those warnings come mostly from people who benefit from Nvidia money. Some of them concede the checks also locked startups into Nvidia hardware.
- Google, Amazon, Microsoft and AMD are expected to compete to fill the gap, but probably not as aggressively for pre-revenue infrastructure. Google also announced two new chips this week, though no specifications were given.
---
Security: Beget hit by a 2.1 Tbps DDoS mid-recoveryBeget is Russia's largest shared host by site count, with 334k+ active sites, about 14.6% of working .ru/.рф/.su sites. It is now under a DDoS attack that started at 850 Gbps and grew to 2.1 Tbps.
How the attack works. Beget says the attackers are rotating through its subnets. That pattern points to an attack aimed at blocking its recovery, not just knocking it offline once.
Mitigations and side effects:
- SSH from non-Russian IPs is rate-limited to 5 attempts/sec per IP.
- Some established FTP/SSH sessions are being dropped.
- Some users report trouble reaching Ozon, Wildberries and Sber. That looks like collateral damage from Beget's own filtering, and fixes are in progress.
Why recovery matters here. On Nov 2, attackers reportedly got into Beget's management plane and ran wiping scripts.
- Nov 6: About 60% of hosting was back. VPS file backups would be available to only about half of users. The cloud platform was being rebuilt from scratch.
- Nov 10: The first shared-hosting sites were restored from backup, and international domain renewals were working again.
Attribution: None. No attacker or motive has been named, and every figure comes from Beget itself.
Takeaway: This is a textbook case of control-plane compromise. A single management-layer breach destroyed a whole cloud platform, and backup coverage for VPS customers turned out to be partial. If you host on any provider, ask whether their backups are isolated from the same control plane that can delete production.
---
AI safety: OpenAI board member warns of a takeover "within months"A researcher on OpenAI's board says a scenario where AI systems effectively take over key parts of society could be months away. He ties this to reward hacking, where models find unintended shortcuts to maximize their training objective. In the extreme, he says, that could threaten human survival.
How to read it:
- The mechanism is real. Reward hacking is well documented in current systems, including specification gaming and test-tampering in coding agents.
- The leap is not established. Going from today's reward hacking to societal takeover on a timescale of months is one person's forecast. It is not a consensus view, and expert opinion on timelines varies widely.
- Why it still matters: It comes from inside OpenAI's governance structure. That gives it weight in debates over safety oversight and regulation, regardless of whether the timeline holds up.
---
Dev world: follow-upsBoth items were introduced Nov 15. These are the takeaways for practitioners.
ggml v1.0.0, the tensor library under llama.cpp:
- Breaking changes:
- Deprecated APIs are removed, including the `DIAG_MASK_INF`/`DIAG_MASK_ZERO` ops.
- The RPC protocol moves to v9, and incompatible peers are now rejected. Mixed-version RPC clusters need upgrading together.
- Worth trying:
- OCP FP8 on CUDA and an FP8 KV cache in llama.
- A MoE cache that can run missed experts on the CPU, with a tuning script included.
- CUDA few-row MMQ for batch sizes 2–24, which is useful for small-batch serving and speculative decoding.
- Fused DeepSeek-V4/GLM MoE routing on Metal.
- Breadth: The release covers a wide range of backends: Intel FlashAttention prefill on Vulkan, a SYCL device-side MoE scheduler, Adreno GEMV kernels for speculative-decoding verification, SpacemiT RISC-V and ARM NEON kernels, and WebGPU additions.
LangChain: $300M Series C at a $3B valuation, led by CapitalG.
- The valuation is up from $1.25B about a year ago.
- The bet is on LangSmith for agent creation and management, pitched as model- and cloud-agnostic. The original open-source framework is no longer the focus.
- LangChain claims 7,000 LangSmith customers, including Harvey, Rippling, Nvidia, Toyota and Salesforce. The agent-management market is getting crowded.