Google, OpenAI, and Meta — AI leaderboard jockeying heats up, the grim forecast of “AI 2027,” and the shadow of looming chip tariffs

Subscribe to policy.ai
Worth Knowing

New Releases from Google, OpenAI, and Meta: The recent pace of model releases hasn’t slowed down in recent weeks, with some of the biggest players in AI — Google, OpenAI, and Meta — releasing new models and features:
  • Google released Gemini 2.5, its most powerful model yet. A reasoning model like OpenAI’s o1 and o3 models, Gemini 2.5 Pro Experimental (the first model from the family to be released), scores highly on a number of benchmarks and is the new top model on the popular LMArena leaderboard. Soon after OpenAI debuted ChatGPT, Google executives reportedly declared “code red” and channeled significant resources into developing and releasing new AI models and tools. More than two years on, those investments are bearing serious fruit: between Gemini 2.5 Pro Experimental, its faster Gemini 2.0 Flash Thinking, Gemini 2.0 Flash, and Gemini 2.0 Flash Lite models, Google now has the best-performing models at a range of price points. But while that’s nothing to scoff at, the company is still having more trouble than it would like turning performance into market share. With OpenAI rumored to be closing in on 1 billion active users, Google staged a shakeup on its Gemini leadership team, elevating Josh Woodward — who oversaw the launch of one of Google’s biggest AI successes, Notebook LM — to lead the Gemini effort.
  • Off the heels of its GPT-4.5 launch last month, OpenAI continued shipping with a number of noteworthy releases, including: vastly improved image generation, improved memory in ChatGPT, the public release of its o3 and o4-mini models, and a new series of GPT-4.1 models for its API users. One of the key takeaways from its slew of releases is that OpenAI, like Google, clearly sees the consumer market as the strategic center of gravity. While the Gemini 2.5 Pro announcement could have stolen some of OpenAI’s thunder — even after the o3 launch, Google still has the top model on the LMArena leaderboard — OpenAI upstaged Google by releasing its improved image generation the very same day (thus flooding the internet with Studio Ghibli-inspired images). But even if they help attract and retain customers, the new models aren’t necessarily better across the board. Some users say ChatGPT now flatters them too much, and OpenAI’s own model card notes that o3 and o4-mini hallucinate more often than the company’s earlier reasoning model, o1.
  • Meta announced its Llama 4 family of models, but the rollout was marred by erratic answers and claims that Meta juiced performance to rank highly on public leaderboards. The new model family comes in three sizes: Scout, which runs on a single Nvidia H100 and boasts an industry-leading 10 million-token context window; Maverick, the mid-range workhorse model; and Behemoth, a two-trillion-parameter “teacher” model that Meta says outperforms GPT-4.5, Claude Sonnet 3.7, and Gemini 2.0 Pro but remains in training for now. Both Scout and Maverick models were released as open-weights models and are available for download through Meta or Hugging Face. At launch, Meta boasted impressive scores for Maverick: on the LMArena leaderboard, it initially had a score of 1417, good enough for second place at the time (now it would be third, behind OpenAI’s o3 model). But it appears that score was for a version of Maverick “optimized for conversationality” — not the one that was released to the public. LMArena has since removed that version of Llama 4 leaderboard and posted a statement on social media saying that “Meta’s interpretation of our policy did not match what we expect from model providers.” The publicly available version of Maverick remains on the leaderboard with a score of 1271, behind other open weights models like DeepSeek’s V3 (score of 1373) and Alibaba’s QwQ-32B (1314).
We’ll probably never see another day-to-day leap like the one from GPT-3.5 to GPT-4, but it’s not for lack of improvement. The biggest AI developers are churning out new models and tools as competition heats up. That competition could be good for consumers, but as OpenAI’s hallucination issues and Meta’s leaderboard shenanigans indicate, it might also mean cut corners.
Huawei’s Power-Hungry AI Supercomputer Is A Serious Nvidia Competitor: The Chinese tech giant Huawei debuted its new AI supercomputing system — the CloudMatrix 384 — that bests Nvidia’s top offering on a raw performance basis. Chip-to-chip, Nvidia’s GB200 chips outperform Huawei’s Ascend series, but as SemiAnalysis detailed in its excellent writeup, the more appropriate comparison is system-to-system: here, Huawei’s system of 384 interconnected 910C chips can deliver an estimated 300 petaflops of computing power, compared to 180 petaflops for Nvidia’s top system, the 72-chip GB200 NVL72. That headline speed comes at the cost of much higher power usage: Huawei’s system uses almost 4 times as much power as Nvidia’s (and still more than double the power on a per-FLOP basis). But as SemiAnalysis points out, that’s a tradeoff China is willing to make. Unlike the United States, which is struggling to find the electricity needed to power its AI boom, China is in the midst of a massive energy infrastructure buildout. As Stratechery’s Ben Thompson pointed out, the size, power consumption, resiliency issues, and (presumed) higher cost mean Nvidia’s system is still the obvious choice in a vacuum, but with access cut off by U.S. export controls, Huawei’s CloudMatrix 384 is an acceptable substitute for AI developers in China. Paired with the recent successes of the Chinese AI lab DeepSeek, Huawei’s launch is a reminder that containment will not be as simple as turning off the tap of high-end AI chips. Future export controls may need to target optics and more kinds of semiconductor manufacturing equipment, or to further tighten prior controls on high-bandwidth memory and AI chips themselves. But even that may not be enough to stop China from developing advanced AI.
AI 2027 — A Dramatic Forecast Sketches the Impacts of Superintelligent AI: A report outlining a hypothetical scenario in which artificial general intelligence (AGI) or artificial superintelligence (ASI) either helps or destroys humanity by the end of the decade has made waves in Silicon Valley and the AI policy space since its publication last month. The report — “AI 2027” — was authored by former OpenAI governance researcher Daniel Kokotajlo, the prominent blogger known as Scott Alexander (author of the popular Slate Star Codex and Astral Codex Ten), and forecasters Thomas Larsen, Eli Lifland, and Romeo Dean and published by the nonprofit AI Futures Project. Set against break‑neck model improvements, escalating U.S.–China competition, and an inadequate regulatory environment, AI 2027 sketches two hypothetical futures: In its “race” ending, fictional U.S. lab “OpenBrain” ignores early signs of deception and, propelled by geopolitical pressure, deploys a misaligned powerful AI system that eradicates humanity. In the “slowdown” ending, the United States slows down AI development just enough to ensure alignment is achieved without being overtaken by China, enabling a safe and prosperous future. ​While the scenario’s authors draw on dozens of tabletop exercises and feedback from more than one hundred experts, their timelines are far from a settled question — in the inaugural post of her new Substack (subscribe now!), CSET’s Helen Toner broke down the range of proposed timelines for achieving AGI. But 2027 is not an unusual prediction — the CEOs of OpenAI, Anthropic, and Google DeepMind have all said they expect human‑level AI before the end of the decade. And some important figures in the world of AI are taking AI 2027 seriously: Turing Award-winner Yoshua Bengio recommended the report, as did Anthropic co‑founder and Import AI author Jack Clark, who called it “the best treatment yet of what ‘living in an exponential’ might look like.” ​
Government Updates

Looming Chip Tariffs Keep AI Sector on Edge Despite 90-Day Pause: After weeks of whiplash — blanket “reciprocal” duties on April 2, a 90-day pause on April 9, China-specific rates raised to 145 percent on April 10 (though President Trump indicated yesterday he may reverse course), and a blanket exemption for most electronic goods on April 11 — tariff chatter has come down a bit from its peak, but AI developers, chipmakers, and the market are still waiting for more shoes to drop. For the AI world, the looming worry is a planned tariff package on chips and the electronics that house them. In February, President Trump said semiconductor tariffs could start around 25 percent and “go very substantially higher over a course of a year.” On April 14, the administration launched its initial investigation — under Section 232 of the Trade Expansion Act of 1962 — that could formally lead to semiconductor tariffs on national security grounds. While those investigations typically take as many as 270 days to complete, Politico reports the administration is hoping to move faster. There are still questions about how such tariffs would be applied, but semiconductor expert Chris Miller thinks they could take the form of “component tariffs” — calculating the duty based on the value of foreign-made components in the entire final product, not just the final product itself. But as Miller points out, the complexity of semiconductor manufacturing means it’s unlikely the tariffs will bring chipmaking back to the United States in time to prevent significant price increases. Higher GPU and server costs would hit AI training budgets just as advertising-heavy giants like Google and Meta face softer consumer spending under broader tariffs. Whether the White House ultimately fine-tunes the plan is anyone’s guess; for now, the sector is stuck in wait-and-see mode.

Ambitious Tech Plan Clashes with Visa Crackdown and Funding Fights: In his first major speech since his Senate confirmation last month, OSTP Director Michael Kratsios promised to “secure America’s preeminence” in AI, quantum, and biotech through lighter regulation and “creatively allocating” federal R&D spending. The vision tracks closely with President Trump’s public March letter to Kratsios, which urged him to “blaze a bold path” and make the United States “the unrivaled world leader in critical and emerging technologies.” Kratsios inherits an innovation ecosystem already under stress from White House policies. Most notably, the administration’s approach to immigration has shown signs of disrupting a vital talent pipeline: a handful of high-profile detentions, denied entries, and revoked (then reinstated) visas involving scientists have some foreign researchers rethinking their U.S. work and travel plans, according to Nature. Even remote collaboration is under the microscope: in March, the White House Office of Management and Budget sent foreign grantees a questionnaire — asking, among other things, whether projects contain “DEI elements” — to gauge alignment with administration priorities. Immigration clamp-downs may restrict the talent pipeline, and funding pressures could hit those who stay: the administration’s high-profile fight with top universities has frozen billions in research grants, and a draft FY 2026 budget proposal circulated earlier this month signaled deep science cuts. Kratsios’ “Golden Age” could prove hard to usher in if the engines of U.S. research — its talent pipeline, collaborative ties, and federal funding base — falter.

U.S. Tightens AI Controls — Blacklists 80 Firms and Blocks Nvidia’s H20: The Trump administration took aim at China’s AI ecosystem with a handful of moves over the last month: first, the Commerce Department added 80 mostly Chinese firms to its Entity List. A few weeks later, the administration imposed an indefinite licensing requirement on Nvidia’s H20 and equally powerful chips. The entity-list expansion sweeps in server maker Nettrix, several Inspur affiliates (a major Intel customer), and the Beijing Academy of AI. The Nettrix addition points to the entity list’s shortcomings — now one of the biggest server companies in China, it was founded only a half decade ago by former executives from Sugon, a server company placed on the entity list in 2019. The department’s move to pull the plug on Nvidia’s China‑specific H20 accelerator could prove more impactful in the long run than entity list whack-a-mole. For Nvidia, the financial hit — a $5.5 billion writedown (plus the broader loss of access to the lucrative Chinese market) — is survivable, but the strategic damage could linger: with Chinese customers set to pivot to Huawei’s competing system (see above for more), one of Nvidia’s most important advantages — the ubiquity of its CUDA software — is harder to defend if one of the world’s largest AI markets is forced onto non‑Nvidia hardware. While there had been some speculation that Nvidia would skirt restrictions after CEO Jensen Huang attended a $1 million-a-head dinner with President Trump at Mar-a-Lago, the administration ultimately seems to have decided the risk wasn’t worth it. With the AI ecosystem moving toward reasoning-heavy models like OpenAI o3 and DeepSeek’s R1, a powerful inference chip like the H20 likely seemed too capable to send to a rival power.

In Translation
CSET’s translations of significant foreign language documents on AI
If you have a foreign-language document related to security and emerging technologies that you’d like translated into English, CSET may be able to help! Click here for details.

What’s New at CSET

ANNUAL REPORT

In 2024, CSET continued to deliver impactful, data-driven analysis at the intersection of emerging technology and security policy. Explore our annual report to discover key research highlights, expert testimony, and new analytical tools — all aimed at shaping informed, strategic decisions around AI and emerging tech.

REPORTS

PUBLICATIONS AND PODCASTS

EMERGING TECHNOLOGY OBSERVATORY

EVENT RECAPS

  • On March 25, CSET hosted an in-depth discussion about AI red-teaming — what it is, how it works in practice, and how to make it more useful in the future. Watch a full recording of the event.

IN THE NEWS

What We’re Reading

Report: The 2025 AI Index, the Stanford Institute for Human-Centered AI (April 2025)

Article: Driven to Self-Reliance: Technological Interdependence and the Chinese Innovation Ecosystem, Yeling Tan, Mark Dallas, Henry Farrell, and Abraham Newman, International Studies Quarterly (March 2025)

Paper: Do Large Language Model Benchmarks Test Reliability?, Joshua Vendrow, Edward Vendrow, Sara Beery, and Aleksander Madry (February 2025)

OSZAR »