Elon Musk's AI 'Supersonic Tsunami' Chart and the New Model Race

When Elon Musk posted a bare blue bar chart to X on August 3 with the caption "AI is a supersonic tsunami," he was doing something he does rarely: making an argument with a picture instead of a prediction. The chart spans July 2023 through July 2026, stays nearly flat for most of that window, and then bends sharply upward the moment 2026 begins. The message was unmistakable even without a label: after two years of incremental gains, something broke out this year. The interesting story is not the metaphor. It is whether the new models shipping out of the major labs actually back it up.

The Chart Everyone Read and Nobody Could Decode

The graphic's biggest weakness became its most-discussed feature within hours. It has an x-axis of dates and rising bars, but no y-axis, no metric, and no source — leaving readers to guess whether it measured compute spending, token volumes, benchmark scores, or pure vibes8. Even Musk's own chatbot, asked to interpret the image, could only shrug in machine form, suggesting the underlying data might be search interest, sales, or user counts while conceding nothing was labeled8.

Tracing the image back revealed an awkward origin: it appears to have come from a near-identical tweet by Paul Graham, who suggested it depicted a company's revenue without naming the company, with the meme annotation "it's so over" added somewhere along the way before Musk reshared it without attribution68. One commentator's blunt verdict captures the epistemic problem — the image supplies a narrative shape but no nails to hang that narrative on, making it useful as a rallying cry and useless as evidence6.

And yet the shape of the chart matches a real, checkable story about the last two years. The flat stretch corresponds to late 2024's genuine trough, when reports said OpenAI's next model, code-named Orion, showed only modest gains over GPT-4 and revived fears that scaling had hit a wall — a wall Sam Altman publicly insisted did not exist24. The upward snap in 2026 corresponds to a year in which the frontier labs have shipped model after model at a pace that has visibly reaccelerated the race. Musk's chart may be unattributed and unlabeled, but it is not dishonest about the direction of the curve.

New Models Gave the Metaphor Teeth

At the moment Musk posted, his own lab had just handed him his best supporting evidence. Grok 4.5, xAI's flagship at the time, had recently topped an independent agent benchmark from Artificial Analysis, beating rival systems on both cost and speed — the kind of outside validation that a company playing catch-up cannot manufacture for itself14. Earlier in the summer, Grok 4.5 had also topped AutomationBench, briefly backing Musk's claim that it matched Claude Opus-class models20.

Musk paired the post with two concrete moves that made clear the tsunami framing was tied to model strategy, not just mood. First, he said SpaceX engineering data would feed Grok's next training run, excluding anything restricted under U.S. arms-export rules, with the stated goal of sharpening the model's reasoning on real-world physical problems where text-only training data falls short14. Second, he reversed his public stance on Anthropic, calling the rival — a lab he had spent months criticizing — "the current industry leader"12. That reversal is arguably the most honest sentence in the whole episode: it concedes that the wave is not solely his.

The Scoreboard Caught Up With the Slogan

What happened in the weeks after the post tells you more about the state of the model race than the chart ever could. Grok 4.7, teased since late July and delayed past a September 12 target, finally shipped on September 211417. The release focused not on chat flair but on long-duration execution, self-verification, and professional knowledge work, with training tasks that take hours to complete18. The independent numbers were solid without being definitive: Grok 4.7 scored 46 on Artificial Analysis's Intelligence Index, a two-point gain over Grok 4.6, good enough to push SpaceXAI into the top four AI labs globally, but still behind Claude Fable 5.1, GPT-6 Astra, and Claude Opus 520. On the AA-Briefcase test of realistic professional work, it jumped 111 Elo points to 1,657, landing just behind Claude Opus 5 and Fable 5.116.

Then came the moment that should anchor any honest reading of the tsunami narrative. On September 28, Musk publicly agreed — with the caveat "for now" — with a ranking chart that placed his own Grok 4.6 and 4.7 three tiers below Anthropic's top-ranked Claude Opus 5.5, with OpenAI's GPT-6 Astra second and Google's Gemini broken out at the bottom1213. This is the same man who, days before Grok 4.7's launch, had promised it would exceed every model on the market20. He has also conceded in interviews that Grok's latest iterations, including the unreleased 2.5-trillion-parameter Grok 4.8, would still not reach par with OpenAI's Astra and Anthropic's Fable class11.

The pattern is consistent: Musk is comfortable losing any given month's leaderboard round as long as he can point to a bigger release on the horizon16. That horizon now has a name. On September 14, he declared that Grok 5 — not 4.8, not the 4.9 he projects at Astra/Fable class — will be xAI's first model to achieve artificial general intelligence, offering no benchmark data, no release date, and no specifications beyond his own hedge: "Maybe better than anything. We shall see"1419.

The Wave Is Real; the Chart Is Marketing

Here is the reading the evidence supports. The acceleration Musk's unlabeled chart gestures at is genuinely visible in the model cadence — every major lab shipped meaningful new frontier systems this year, benchmark gains of 100-plus Elo points between single generations are now routine, and a 111-point jump between Grok 4.6 and 4.7 is exactly the kind of step change that would make an old flat curve bend1618. The costs of that pace are equally visible: Grok 4.7 burns roughly 81,000 output tokens per Intelligence Index task, more than double its predecessor and triple GPT-6 Astra, meaning per-task bills rise even when per-token prices hold20.

Wall Street, notably, has stopped arguing about whether the wave exists. Investors have shifted from debating AI's transformative potential to scrutinizing who pays for the hundreds of billions in data center, chip, and networking buildout — with hyperscaler capital expenditure now treated as the single most important indicator for the entire AI supply chain, and markets increasingly punishing companies for the spending itself9. That is a very different conversation from 2024's "is scaling over?" panic, and it is itself evidence that the curve moved.

The expert split on where the wave lands remains wide. Musk, echoing his Davos remarks, says AGI could surpass a single person's intelligence before 2026 ends, and separately projects AI exceeding all of humanity's combined intelligence within about five years, framing it as both an age of extreme material abundance and an existential risk124. Geoffrey Hinton puts broad AGI as much as two decades out24, while Ben Goertzel of the Artificial Superintelligence Alliance argues advanced systems already rival state-level capabilities2. Musk has been making the tsunami case for a while — he told Joe Rogan last year that digitally centric desk jobs would be taken over "like lightning" while physical work endures longer, promising "universal high income" alongside "a lot of trauma and disruption"10.

Where This Leaves the Tsunami Thesis

Two months on, the fairest verdict is split down the middle. Against Musk: the chart is borrowed, unlabeled, and unattributable to any dataset; his own flagship trails the leaders on both the tiered chart he endorsed and independent indices; and his boldest claim — Grok 5 as the first AGI — rests entirely on assertion681214. For Musk: the labs are shipping at a velocity that has visibly reaccelerated since the November 2024 trough, his company jumped 111 Elo points in a single generation and cracked the independent top four, and even his grudging naming of Anthropic as the leader implicitly confirms that the frontier is moving faster than almost anyone predicted two years ago1620.

The supersonic tsunami, in other words, is better understood as a competitive observation dressed up as a prophecy. The models broke out in 2026. Who ends up surfing the wave — and who gets buried under the capex required to keep generating it — is the question the rest of the year will actually answer.