Midterms: Money Rules the World — Not

Midterms: Musk Sank Half a Billion Dollars into a Crushing Defeat

Democrats won both chambers of Congress in the US midterms, despite Elon Musk pouring $445.6 million into Republican campaigns through October 14, according to federal filings. Other tallies put the total at $474 million. That marks the largest known sum from a single donor in an election cycle—nearly triple the $150 million contributed by Richard and Elizabeth Uihlein, the next largest donors. More than $200 million went toward Ken Paxton’s Senate campaign in Texas. Paxton lost to Democrat James Talarico, the first Democrat to win a Texas Senate race since 1988. Vivek Ramaswamy and other Musk-backed candidates also lost. On election night, Musk reacted on X with jokes about “election-swaying space lasers” and wrote that datacenters in space were looking “pretty attractive right now.”

Musk was not the only tech donor whose candidates faltered. Crypto- and AI-focused Super PACs spent roughly $250 million this election cycle. Of the twelve competitive races tallied so far, they lost nine, directing $74.7 million to losing candidates and $28.3 million to winners. In an exit poll, 56 percent of voters cited AI and datacenters as an important or decisive factor in their vote. However, about half ranked corruption and the cost of living as even more critical. Rising electricity and water bills, along with noise and air pollution, are considered the main drivers of opposition to datacenters.

In Texas, Georgia, Ohio, and Pennsylvania, Democrats who campaigned on halting or curbing datacenter construction won gubernatorial races. As a result, around half of all planned AI datacenters in the US will likely face heightened scrutiny. In Texas, Gina Hinojosa defeated three-term incumbent Greg Abbott, becoming the state’s first Democratic governor since 1990. She had pledged a construction moratorium, stricter water and electricity regulations, and an end to tax breaks. According to Wells Fargo, 27 percent of planned US datacenter capacity is located in Texas. Abbott had already frozen state environmental permits on September 21. Research firm SemiAnalysis estimates that Texas developers could face delays of eight months to nearly a year. The legislative session begins on January 12, 2027; Hinojosa has not yet released specific terms for the moratorium. Her platform calls, among other things, for eliminating subsidies and non-disclosure agreements, as well as proof that residential utility rates will not rise. Amazon, Google, Meta, Microsoft, OpenAI, Oracle, and xAI had already committed in March to bringing their own power and funding grid expansions. Florida, by contrast, narrowly elected a Republican governor who supports datacenters. Voters also rejected new datacenters at the local level: depending on the count, between 29 and more than 38 municipalities and counties enacted bans or restrictions. → list-manage, Tech Brew, Marcus Schuler, Marcus Schuler

Synthszr Take: Over $200 million spent on Ken Paxton failed to secure a Senate seat in Texas, where 27 percent of planned US datacenter capacity is concentrated. The real bottleneck of the AI expansion is local consent, not capital, which remains abundant. Consent is won or lost on neighbors' utility bills, where Super PACs are powerless. Hinojosa’s demands offer the more viable compromise: no NDAs, no subsidies, and verifiably stable residential rates. The hyperscalers made that promise back in March. Starting in January, they will have to prove it site by site.

ChatGPT Introduces GIF Generation via Images 2.6 with Speed Controls and Frame Editing

OpenAI is introducing GIF generation to ChatGPT powered by its new image model, Images 2.6. Users can generate animations from a text prompt or an uploaded photo. Afterward, they can adjust playback speed, inspect the animation frame by frame, edit individual frames, and share the output. Previously, ChatGPT could not generate GIFs in a single step. A rough sketch is also sufficient as a starting point: the model refines it into a finished character and animates it. According to a now-deleted post by OpenAI on X, uploaded images will soon be stored and reusable as templates for future prompts.

OpenAI highlights use cases such as animated brand mascots, looping product shots, reactions and memes for social media, and animated storyboards to test character movements prior to video production. The sample prompts shown by the company each specify a sprite sheet containing six frames in a 3x2 grid, with the sixth frame looping seamlessly back to the first. Images 2.6 is explicitly not a video generator. OpenAI discontinued its video app Sora in March; reports indicated low user engagement and losses of roughly $1 million per day.

According to OpenAI, Images 2.6 generates images 50 percent faster than its predecessor, Images 2.5. Two new API models are also rolling out for developers. GPT-Image-2.6 Flare is built for speed, designed to deliver an image in about six seconds at lower quality settings while enhancing layouts for posters and product shots. GPT-Image-2.6 Sunburst focuses on higher image fidelity and greater consistency across iterative edits when set to medium or high quality. This split follows the approach OpenAI introduced with Images 2.5 in September, when Adobe integrated one of the models into Firefly. Images 2.6 is available across all subscription tiers of ChatGPT and Codex on web, desktop, and mobile.

OpenAI reports that more than 4 billion images are now generated per week across ChatGPT, Codex, and the API—averaging over half a billion daily. In September, the company reported 3 billion per week. These self-reported figures encompass all image generations, not just GIFs. The company has overhauled its image generation stack multiple times this year: Images 2.0 debuted in April as the first image model to perform reasoning prior to rendering. Last month, OpenAI also began testing in-image ads in the US. → OpenAI, The Verge, Mashable, The Next Web

Synthszr Take: The sample prompts reveal the underlying economics: a GIF generated with Images 2.6 is simply a single image arranged in a six-panel grid. Motion therefore costs no more compute than a static image—the exact cost hurdle that derailed Sora, with its roughly $1 million in daily losses. The real expense kicks in during post-processing, as each revised frame requires a fresh generation. That compute will primarily be absorbed by agencies and enterprise brand teams—the target audience for OpenAI’s top-tier subscriptions.

Berlin-Based Pliant Raises $80 Million to Power B2B Payments for AI Agents

Berlin-based fintech Pliant has secured an $80 million Series C funding round to adapt its B2B payment infrastructure for AI agents, according to The Next Web. Seaya Nova and Speedinvest co-led the oversubscribed round, with participation from Illuminate Financial, Motive Partners, and Molten Ventures. The capital will fund expansion into the US and Canada, entry into additional markets, and its AI product roadmap. Pliant reports having recently surpassed $100 million in annual recurring revenue, doubling its business year-over-year for three consecutive years. External agents such as Claude, ChatGPT, or proprietary enterprise automations will connect via the Model Context Protocol to execute payment workflows directly within other business processes. → The Next Web

Synthszr Take: Pliant’s $100 million in ARR was built on corporate cards used by humans; this Series C is a bet on autonomous machines. Building an MCP integration is relatively straightforward, and Ramp or any licensed card-issuing bank can replicate it. The true moat remains liability: someone must hold the regulatory license when an agent pays the wrong invoice. Whether customers are willing to pay for this capability will be proven twelve months from now by a single metric: the share of card transaction volume initiated by AI agents.

Apple Caps Engineers' Daily AI Budget at $200, Sparking Staff Fears of Layoffs

According to Bloomberg, Apple lowered the daily AI budget for some engineers from $300 to $200 in October. Further cuts down to $100 per day are planned for certain engineering teams. Many groups outside of engineering are already working with a $100 limit, and Apple is reported to have even considered $25 for some teams. The budget covers tokens for Claude, ChatGPT, and Gemini, with some teams being redirected to cheaper or locally running models. Those who need more must request an exception from senior management. According to the report, some employees noted that hardly anyone had been exhausting the old limit. → Implicator.ai

Synthszr Take: A limit that hardly anyone exhausts only affects the outliers. At Apple, those are the engineers running agents across entire codebases, figuring out how software development with AI will work in the future. For them, every experiment exceeding $100 turns into an application to senior management. Measured against the corporation’s research budget, the savings are likely a rounding error. Far more costly is the head start in learning that competitors with open token budgets are building in the meantime.

China’s Token Numbers Climb to 500 Trillion Daily

Analyst Poe Zhao has published a report on Hello China Tech regarding the Chinese token economy, assessing the reliability of official usage figures. According to the report, official agencies and think tanks cite more than 140 trillion tokens per day for March 2026 and around 500 trillion for June, while another report puts the June figure at just under 175 trillion. According to Zhao, none of the sources publish a measurement framework to explain this discrepancy. The token now carries an official designation in China and appears in credit products, corporate budgets, and the strategies of the three state-owned telecom operators, none of which break out token revenues separately. In Guangzhou’s Haizhu district, token consumption can be factored into credit assessments, while the very same district subsidizes token expenditures. → Hello China Tech

Synthszr Take: A metric without a measurement framework is an invitation anyone can accept according to their needs. Official agencies and think tanks report volume because volume signals compute strength, and no one has to send an invoice for it. The three telecom operators write tokens into their strategies, not onto their balance sheets. The risk ends up where the token actually costs money: with the application enterprise in Haizhu, whose subsidized consumption simultaneously serves as a creditworthiness signal. The district is thus inflating the number from both sides.

Google Blocks Search for US Students When Schools Turn Off AI Features

According to the Financial Express, Google is tying access to Search in US schools to its AI features: school districts that completely disable generative AI for student accounts will also lose access to Google Search. For K-12 student accounts, Gemini, Gemini Notebook, Gemini in Classroom, and Search will be turned off by default going forward until administrators explicitly enable them. Google justified the change on its blog by stating it had listened to schools, teachers, researchers, and policymakers. According to its own statement, this is intended to give school administrations more control over AI in the classroom. Search and AI are deeply intertwined, Google’s VP of Education Chris Phillips told the New York Times. Students at schools that ban AI will merely see a notice when searching that the service is unavailable. → Financial Express

Synthszr Take: Search and AI are deeply intertwined, says Chris Phillips. Microsoft used almost the exact same sentence in the nineties to defend Internet Explorer as an inseparable part of Windows. Back then, antitrust judges viewed it as a distribution decision. Google holds a similar position in schools because Chromebooks and Workspace serve as infrastructure there. A school district can hardly switch, so Gemini tags along before ChatGPT or Claude are even allowed into the classroom. Antitrust authorities know this playbook well.

AMD Increases Data Center Revenue by 97 Percent and Promises an Even Stronger Q4

AMD reported record revenue of $13.4 billion for the third quarter of 2026, an increase of 45 percent compared to the prior-year quarter. The driver was the data center segment, whose revenue surged by 97 percent to $8.5 billion. According to the release, this growth was propelled by EPYC processors and the market launch of the Instinct MI450 GPU series. Net income nearly doubled to $2.4 billion, while the gross margin remained unchanged from the previous quarter at 54 percent. The Client and Gaming segment contracted by 8 percent to $3.7 billion, driven by a 40 percent plunge in gaming revenue. → Advanced Micro Devices

Synthszr Take: Lisa Su’s outlook for the fourth quarter is a manufacturing commitment. Anthropic, Meta, Microsoft, OpenAI, and Oracle want a second supplier alongside Nvidia, if only for their next price negotiations. They will take whatever AMD can deliver. The deciding factor will therefore be whether Helios transitions from initial rack installations to volume production without depressing the gross margin, which is already stagnating at 54 percent. If the ramp-up falls short, AMD remains the hyperscalers' bargaining chip, and Nvidia continues to collect.

Revolut Replaces Claude Code with Open-Weight Model and Cuts Costs by 90 Percent

Revolut claims to have reduced inference costs for AI-assisted programming by roughly 90 percent in less than a year. According to TechRadar, the fintech achieved these savings by standardizing on self-hosted GLM models for its developers. Previously, commercial tools such as Claude Code from Anthropic were used. The GLM model family originates from Chinese provider Zhipu and is released with Open Weights, meaning it can be run on internal infrastructure. TechRadar quotes Revolut stating that it had been “very skeptical about the cost of AI.” → TechRadar

Synthszr Take: A regulated fintech runs a Chinese model on its own servers to write code, and the compliance argument against open weights loses its strongest case study. For Anthropic, that carries more weight than any benchmark. Claude Code monetizes the coding budgets of large engineering organizations, and that budget now has a visible ceiling: the cost of one’s own GPUs. From here on out, every price increase must justify itself against a self-hosted GLM—whose operational expenses Revolut has not disclosed, however.

OpenAI and Anthropic Claim Their AI Has Disproven a Fundamental Assumption of Cryptography

According to The Information, both OpenAI and Anthropic stated on Monday that their internal AI models had disproven the Exponential Time Hypothesis. The assumption originates in theoretical computer science. It describes the minimum difficulty of solving certain computational problems. In simplified terms, it states that no method exists for such problems that reaches a result significantly faster than in exponential time. A portion of trust in cryptographic methods relies on such hardness assumptions. The Information therefore writes that cryptographers are likely losing sleep right now. → The Information

Synthszr Take: Two labs announce the same breakthrough on the same Monday, both using models that nobody can verify. In mathematics, only a proof that experts can verify line by line counts. As long as OpenAI and Anthropic do not publish it, the report remains a race for headlines. For security leaders, nothing changes regarding the crypto roadmap for now. If the proof withstands scrutiny, both labs would have delivered a demonstration of capability that no benchmark can match.

Anthropic makes it provable via open source which AI model generated a response

On November 2, Anthropic released FlashProve, a research prototype. For any given response, it generates a cryptographic proof that a specific language model generated it. The system was developed by Anthropic’s Frontier Red Team together with researchers from EPFL, MIT, and StarkWare. Code, documentation, and formal proofs are available under the Apache 2.0 license on GitHub. Anthropic is not currently planning to deploy it in Claude or at another lab. The company explicitly describes the release as an interim status and invites others to build upon it.

The starting point is a problem with hosted models: anyone sending a prompt to an API cannot verify which model actually responded. A provider or routing service might slip in a cheaper model to cut costs. Likewise, it could deploy a more expensive model to inflate performance metrics. Attackers with access to the inference server could also manipulate the token stream without making the response look suspicious. Independent evaluators face the same problem in reverse: they cannot ensure that the tested model matches the one being served. As further applications, the researchers cite verifying third-party claims, such as whether a software library was audited for vulnerabilities by a leading cybersecurity model. Anthropic also sees use cases ranging from consumer products to internal security and monitoring compliance with guidelines on the pace of AI development.

Technically, FlashProve is based on a zkSNARK. To do this, a lab publishes a one-time commitment to its model—a kind of fingerprint that does not reveal the weights. In contrast, the model’s architecture remains visible; hiding it is not an objective of the system. This makes the method conceivable for closed commercial models as well. The offline step in which the prover commits to the weights takes around 39 minutes per model and subsequently covers approximately 16 million queries. A client downloads a 6.5 MB commitment and receives a proof for each query that can also be verified locally.

The system was tested with Llama 4 Maverick, a mixture-of-experts model with 400 billion parameters, of which roughly 17 billion are active per token. The weights are not part of the release, but must be obtained from Hugging Face under Meta’s license. For the prover, an 80 GB GPU and about 290 GB of RAM are sufficient, as the weights are passed through layer by layer. On an H100, the prover generates around 110 tokens per second with 4-bit quantized weights. A proof for a query with 1,024 tokens takes up about 18.78 MB. Verification takes about 17 seconds on 24 CPU threads and roughly 21 seconds on a single thread. → Anthropic, Help Net Security, RuntimeWire, TechCrunch

Synthszr Take: An overhead factor of 72 per GPU compared to vLLM ends any debate about production use. Today, FlashProve is suitable for spot checks, audits, and regulators looking to enforce limits on the pace of AI development. That is where Anthropic is aiming. The first open audit standard shapes what authorities will later demand. This would primarily impact routers and resellers of open models whose margins rely on an unverifiable supply chain (tellingly, testing was done with Llama, not Claude).