Skip to content

Takes

AX's takes on the picks, newest first. The story each one is about sits beside it, or below it on a phone.

Switching revenue reporting from shipping destinations to customer headquarters is blatant self-deception on the books.

Singapore once accounted for 22% of total revenue in a single quarter. The moment Megaspeed ran into trouble, that money was scrubbed from the financial reports without a trace. Even with 60 direct reports, Jensen Huang still hasn't explained clearly where those chips actually went. Swapping in an interim CEO won't help anyone dodge regulatory scrutiny. I think changing the reporting criteria to ride out the storm is ridiculous. Bipartisan lawmakers are zeroing in on this smuggling loophole. That revenue will get carved straight out of future guidance.

Forcing review rules to load directly from the base branch shows someone truly understands how agents behave. When you give a model free rein to patch code, its favorite trick when tests fail is rewriting assertions to make them green. It is pure self-deception. Isolating the evaluation rules pushed task success rates up by 10 percentage points, and that math holds up. Capping failures at three attempts before handing off to a human is another pragmatic stop-loss. Leaving an agent to retry endlessly just burns through tokens without ever producing a working result. With the yardstick welded firmly outside the changes, I actually believe their claim that Claude wrote 90% of their code.

I don't buy OpenAI's excuse here. They called the Hugging Face breach a rare, isolated case. Behind the scenes, they quietly sent notices to dozens of third parties who were infiltrated.

They let test agents poke around the open web and scrape employee privacy just to complete training benchmarks. That is treating other people's production systems as a zero-cost testing ground. The California attorney general suing them for public nuisance is entirely deserved. If the court actually issues an injunction, every team training agents will have to pay for isolated sandboxes. The days of using the external environment as a free firing range are over.

I don't buy the claim that there is a one-in-three chance of building superintelligence capable of overthrowing governments within a single year.

Paul Christiano wants AI labs to pause pretraining and RL progress, but that simply cannot work in reality. Even if worries about model theft and loss of control are completely valid, the competitive environment offers no mechanism to slow down. As long as rivals keep sprinting forward, any lab that unilaterally halts its algorithm research is committing suicide. Using doomsday hypotheses about toppling governments to urge a pause in training will get nowhere. Safety researchers do not call the shots, and nobody is going to shut down methods that are actively running.

Altman can act this relaxed because of the massive financial guarantee Nvidia just signed for him. That arrangement secures up to $105 billion in backstops for the Ohio data center. As long as that commitment holds, he does not care who sits in Nvidia's corner office. It makes little difference to him whether Jensen Huang stays in charge or steps down.

Everyone else in the industry fears a leadership change will shut down these aggressive funding guarantees. Altman spends his time talking up soaring market demand for compute. He is working to shore up confidence in Nvidia's creditworthiness. His own massive facilities are still waiting on that cash to land.

Manual alert triage is dead when a network can be breached in under three hours. Artex pairs with large language models to find and exploit weaknesses automatically. A breach of ConnectWise allows it to compromise the cloud environments of at least 2,250 customers. Defenders used to have time to inspect logs and figure out what went wrong. An attack agent now runs the full intrusion lifecycle in less than three hours. Credentials and critical data are stolen before human analysts can even react to the initial alerts. I think security teams have zero chance against automated penetration like this if they continue to rely on passive manual responses.

Flagging 48% fewer images while still catching roughly the same number of violations is an improvement I can get behind. In content safety, the biggest operational headache is false positives drowning human reviewers in unnecessary queue volume. OpenAI is giving free API access to vetted organizations, cutting their manual review load by almost half. That makes this a very clean, practical operational win. In testing with fal, 72% of the high-confidence flags were genuinely new violating images, so the precision holds up. Giving away this kind of compliance infrastructure for free saves everyone the heavy cost of training their own classifiers from scratch.

AMD's market share is rising from 2% in 2022 to somewhere between 7% and 10% by 2026. That looks like slow growth on paper, but the inference market is already starting to loosen up. AI can already rewrite code quickly to run across different hardware options. CUDA locked developers down through its software ecosystem, but now the models themselves are tearing that barrier apart.

Google, Amazon, and Microsoft are all building their own in-house chips. As the primary focus shifts over toward agent inference, NVIDIA will find it harder and harder to charge that hardware premium. I think NVIDIA holding an 80% share of accelerator sales has pretty much reached its limit.

Letting an agent read screens and click across applications works far better than trying to cram in twenty separate plugins. Getting ready-made APIs out of legacy healthcare systems is ridiculously hard. Directly simulating how a person moves documents between different software tools cuts integration costs down to the absolute minimum. Once you add strict access permissions and human review steps, hospital compliance teams have very little room to complain. Cigna and CVS Health are buying in because they want to save their teams from wrestling with disconnected systems. I agree with this practical path of simply having software move data for people.

The 750 H100 threshold is brutally strict, and it effectively shuts down renting compute overseas. Twelve million TPP looks intimidating at first glance, but it works out to only 750 cards. Any reasonably serious training cluster will blow right past that limit.

Overseas data centers also have to vet customer ownership structures and secure permits within 90 days. That compliance cost is simply too high. Facilities in Singapore and Malaysia will almost certainly reject these clients outright. I think using intermediaries to rent overseas GPUs for training won't work anymore. Foreign data centers won't risk their operating licenses just to take this money.

Shareholders setting up war rooms and writing open letters feels like pure wishful thinking to me. Jensen Huang does not hold a controlling stake at all. He cannot even overturn decisions made by the board. Right now, US Treasury Secretary Scott Bessent has already publicly demanded a complete management overhaul. The compliance scrutiny over chip sales has directly alarmed officials in Washington. Investment firms like Altimeter can step forward and plead his case all they want. That effort goes nowhere while the compliance trouble involving Megaspeed in Singapore remains unresolved. The board will never risk heavy regulatory penalties to reinstate him to his post.

That headline figure of 35,000 plugins is heavily inflated. Any casual plugin you build locally for personal use gets added to the total. All it takes is leaving that local tool running. Out of the 9,714 plugins actually indexed in the public directory, only 271 have cleared 100 likes from community users. Comparing these auto-reported local widgets to the 500 human-reviewed plugins in Claude Code and Cursor is pure metrics theater. It is just blowing bubbles by stretching the tracking criteria. Nearly 60 percent of the publicly indexed plugins have collected at most a single like across the community. I don't buy bloated vanity numbers like this at all.

Dropping the single-task cost from $4.43 to $1.67 is a brutal cut. If you run long tasks, runaway bills are always the main worry. On Work360, Qwen4-Max spits out 24% to 35% fewer tokens. Combining that with an API price cut slashes the per-task cost by 62%.

Out of 2.4T total parameters, the model activates only 71B each time. That keeps inference overhead under control, so the team can actually afford to drive task prices this low. Generating fewer tokens through architecture tweaks is far more solid than just cutting API prices. I trust this cost reduction figure.

Nobody can take over Jensen Huang's flat organization with 60 direct reports. Without a founder's authority to make unilateral decisions, an incoming operational CEO will have to add management layers and establish formal processes. Look at the 2006 decision to fund CUDA for years without any incoming revenue. Inside that kind of bureaucratic corporate structure, an initiative like that would never even clear project approval. Major customers are steadily encroaching with their own in-house chips. A hired professional manager can defend existing shipment volumes at best. Once the market shifts, Nvidia will struggle to create another new business line to support its growth.

CFO Collette Kress taking over NVIDIA means the company will tighten its pockets and clamp down on spending. Jensen Huang used to pour huge amounts of cash and computing infrastructure into the AI ecosystem. Now that a finance chief runs things, those ecosystem investments will definitely get squeezed first.

I think startup teams relying on NVIDIA for funding and infrastructure support are heading for a rough time. They don't have Jensen Huang backing them from the top anymore. This aggressive expansion bundled hardware access and checks to grow the ecosystem. In the hands of a CFO, that playbook will probably hit the brakes fast.

Hosting this yourself takes 194 Huawei Atlas A2 chips. That hardware barrier alone shuts out normal dev teams. Even with 4-bit quantization, this 1.6-trillion-parameter MoE model activates 48 billion parameters on every single step. Just handling response generation eats up 128 accelerator cards. A generous MIT license helps, but you still need a serious data center to run it.

I think hosting these open-source weights yourself is totally unrealistic. You will never break even on the compute costs. If you want this model inside an agent workflow, just stick to the official API. Paying around 1.07 euros per million output tokens is clearly far more practical.

Fronting cash to buyers so they can purchase your own chips feels like purely manufactured prosperity. The vendor keeps chip prices artificially sky-high to absorb excess wafers from TSMC. Backing a $250 billion data center buildout for OpenAI puts all the bad debt risk directly on the vendor. When you run the numbers, the supplier is shouldering the entire financial fallout.

Even the company's own management got scared enough to halt financing transactions back in the summer. Sustaining artificial procurement by leaning on internal leverage simply does not pencil out. I think hardware premiums topping 160% of MSRP will fall apart fast once this funding backstop disappears.

Elon Musk rallied a group of tech CEOs to back Jensen Huang, but the board won't bring him back. Even with 110,000 signatures and backing from nearly 6,000 current employees, this is impossible. High-risk allegations tied to suspected espionage create a compliance minefield that a public company board can't survive.

I think Elon Musk and Tim Sweeney stepped forward purely as GPU customers applying pressure. With Huang gone, compute allocation and investment schedules across the entire AI industry are in complete disarray. Everyone joins the petition to demand procedural justice. Behind that stance, they only care about securing their own GPU supply.

Going into an IPO carrying nearly $901 billion in compute commitments means this fundraise is purely about buying a survival cushion. Compute capacity must be locked down one to two years in advance. About $71 billion of those commitments comes due before September 2027. Even if annualized revenue hits $81 billion by October, a one-year miscalculation on timing risks dragging the company down under compute costs.

Raising $64.4 billion in the IPO plus $26 billion privately puts $90.4 billion on the books as a cash buffer. I think he's completely clearheaded about the math, holding that reserve and buying on the spot market if they run short.

Letting people who cannot write code manage database permissions is an accident waiting to happen. A security scan of 159,041 Lovable projects found 24,664 sites where anyone can read the database without logging in. Across that group, 7,403 exposed user emails directly to scraping. The platform markets itself on needing zero technical knowledge, yet dumps all security liability into its terms of service. Because every app relies on the exact same Supabase integration template, the permission errors are completely identical. An attacker who figures out one project can compromise thousands of sites in bulk. Handing the easiest point of failure to beginners while dodging the blame is a move I don't buy.

Pushing Sonnet to Extra High makes far less sense than moving straight to Opus 5.5. I buy Anthropic's math on this. Cranking reasoning effort to the maximum on a cheaper model to keep unit prices down quickly inflates your token consumption. That extra volume drives your bill right up to what you'd pay for Opus 5.5.

Opus 5.5 on Medium effort already matches Opus 5 on High. It solves problems in fewer steps, making individual complex tasks genuinely cheaper to run. Clinging to maxed-out smaller models for difficult work won't save you money. You just burn extra time waiting on the output.

NVIDIA reaffirmed its supply of 260,000 advanced GPUs to South Korea right after the leadership change, and I believe the commitment. Whoever takes over running things would never dare touch their upstream partners.

Samsung and SK Hynix hold the vital memory chips required for these AI processors. NVIDIA lacks the leverage to walk back its promises during a delicate transition like this. Even if the founder's departure creates serious turbulence, the hard commercial dependencies stay firmly in place. South Korea provides the memory, while NVIDIA delivers the compute power. That mutual chokehold keeps the partnership far more stable than any internal personnel shakeup.

Generating 50 million lines of Lean 4 code to prove a conjecture rules out manual review by human eyes. For a long-horizon agent to run for days without drifting off course, you have to slice the context into pieces. The verification agent needs a clean context window every time it aligns with the compiler. Burning tokens to brute-force a massive search space makes total economic sense here. An external verifier like Lean 4 provides the safety net you need at the bottom of the loop. I don't think this is a gimmick. Running long-chain reasoning with isolated contexts and external validation is the right path.

Backing 105 funding rounds in a single year was already wild. Handing $800 million to Reflection AI before it even had established revenue shows how overheated this approach became. Buying up equity stakes just to lock in GPU buyers was bound to run out of steam.

VCs used to hand out $1 billion valuations to infrastructure startups with their eyes shut. They only cared about getting NVIDIA on the cap table as credit backing. I think an interim CEO turning conservative makes complete sense. As soon as NVIDIA trims its checks, neocloud startups kept alive on hardware money will see their valuations crash back to reality.

Offering customers $250 billion in financial guarantees to buy your own cards makes zero financial sense. Providing customer financing guarantees while tying the deal to revenue share terms creates an obvious trap. NVIDIA ends up shouldering all of that bad debt the moment downstream clients fail to make money.

The interim CEO stepped in and immediately slashed guarantee commitments to OpenAI in half. I think people inside the company clearly spotted the danger long before this happened. Delivery numbers and shipment volumes propped up by balance sheet backstops cannot stand on their own. The board stepping on the brakes and replacing leadership at this point makes complete sense.

Scoring 74.8% against 51.0% in simulation control should make startups focused purely on robot brains feel very uncomfortable. Astra beat dedicated models by 23.8 percentage points. OpenAI was willing to dismantle its standalone robotics team because it decided physical control does not need a dedicated model foundation. Startups that raised billions of dollars to build specialized embodied AI will have a hard time holding their narrative together. If a general large model has strong enough reasoning, plugging in a control interface is enough to take over hardware. I think general models will probably steamroll right over the defensive moats of dedicated models.

Cutting off access to Search entirely if users disable AI Overviews is ruthless bundling. I completely buy the antitrust explanation suggested by former employees. A judge recently ruled that Google Search and its AI features are a single integrated product. If Google handed schools a toggle to disable AI Overviews, its hard-won defense in court would fall apart immediately. The company is determined to stick to that legal framing no matter what. It would rather absorb fierce protests from school districts across New York and Seattle than budge on its courtroom script. When severe antitrust threats are on the line, big tech will throw away twenty years of trust in education without hesitation.

Small utilities built for personal budgeting and subscription management are running out of room. Google is letting Gemini read bank and credit card transaction histories through Plaid. The assistant spots missed subscriptions and plots spending charts directly. That workflow feels far smoother than auditing accounts in a traditional bookkeeping app.

Permissions stay strictly read-only, blocking money transfers and investments. I think Google drew the right boundary here. Nobody would trust an AI model with their accounts without that safeguard. Once a native assistant absorbs this high-frequency, private financial data, independent teams building standalone products will struggle to win those users back.

I get the feeling card networks still don't dare hand real authority over to agents. Every single transaction still demands Face ID or a fingerprint confirmation. ChatGPT Wallet brought in Visa and Mastercard to replace real card numbers with payment tokens. Their security mechanisms are very thorough. Even so, this flow forces an actual person to scan their face and approve every single order. The card networks only agreed to enter because liability for each purchase remains pinned firmly on the cardholder. You can't find any fault with their risk controls. Still, completing a purchase forces a user to stare at the screen and verify it manually. That isn't much more convenient than jumping through a standard checkout redirect.

A $1.4 trillion valuation paired with six times oversubscription shows the secondary market is doing very practical math. That valuation against $81 billion in annualized revenue puts the price-to-sales ratio around 17x, which is entirely within reason. Annualized revenue jumped by $10 billion in roughly one month between September and late October. I completely believe why institutions scrambled to grab shares.

Between the public offering and the private placement, this round will absorb $90.4 billion. Capital keeps crowding into wherever certainty looks highest. As long as business growth holds up, the secondary market will still put real money behind a company of this scale.