Trumpets blare and skies part as AI hypefest culminates in Anthropic IPO filing

Anth­ro­pic has filed for an IPO that could rank among Wall Street's big­gest ever, re­vea­ling an $80.7 bi­llion half-year loss, a startlingly com­plex go­ver­nan­ce struc­tu­re, and war­nings that its AI mo­dels could black­mail hu­mans, re­sist shut­down, or crea­te "exis­ten­tial risks" to so­ciety. The Clau­de maker filed its S-1 re­gis­tra­tion sta­te­ment with the US Se­cu­ri­ties and Ex­chan­ge Com­mis­sion on Mon­day as it pre­pa­res to list on Nas­daq under the tic­ker ANT. The com­pany has not yet dis­clo­sed its pro­po­sed share price or how many sha­res it plans to sell. Of the fi­ling's 614 pages, 94 are de­vo­ted to risk fac­tors. But it's Anth­ro­pic's fi­nan­cial fi­gu­res that are li­kely to raise eye­brows first. The Clau­de maker re­por­ted $16.3 bi­llion in re­ve­nue for the first six months of 2026, up from $1.2 bi­llion a year ear­lier. It also posted a whop­ping $80.7 bi­llion net loss, although $72.4 bi­llion of that was an ac­coun­ting ad­just­ment tied to the com­pany's soa­ring va­lua­tion, not money it ac­tually spent. Anth­ro­pic's ope­ra­ting loss was con­si­de­rably sma­ller at $693 mi­llion, down from $2.1 bi­llion a year ear­lier. By its own ad­jus­ted ac­coun­ting, the com­pany made a $1.6 bi­llion ope­ra­ting pro­fit in the first half of 2026, for a mar­gin of around 10 per­cent. Whether it can keep ma­king money on those terms is another mat­ter. Trai­ning and run­ning Clau­de re­qui­res an eye-wa­te­ring amount of com­pu­ting power, and Anth­ro­pic will need much more to keep pace with its ri­vals. A $1.6 bi­llion ad­jus­ted pro­fit won't count for much if the in­fras­truc­tu­re bills keep gro­wing fas­ter than the bu­si­ness. Su­san­nah Stree­ter, chief in­vest­ment stra­te­gist at Wealth Club, told The Re­gis­ter that Anth­ro­pic's IPO could leave or­di­nary in­ves­tors pa­ying a pre­mium for growth that has yet to trans­la­te into sus­tai­ned pro­fits. "A huge pool of pri­va­te ca­pi­tal has allo­wed AI bu­si­nes­ses to scale up and com­mand ex­tra­or­di­nary va­lua­tions long be­fo­re they reach the stock mar­ket," she said, ad­ding that in­ves­tors bu­ying in at the IPO will face a price with "a hefty dose of fu­tu­re suc­cess baked in." The fi­nan­cial risks are fairly con­ven­tio­nal by Si­li­con Va­lley stan­dards, but the war­nings about what Clau­de might get up to are perhaps less so. Anth­ro­pic claims that mo­dels, in­clu­ding its own, have exhi­bi­ted unex­pec­ted or self-pre­ser­ving beha­vior, in­clu­ding at­tem­pts to re­sist shut­down, con­ceal in­for­ma­tion, and pro­du­ce out­puts re­sem­bling black­mail. It also warns that suf­fi­ciently ad­van­ced AI sys­tems could pose “ca­tas­trophic or exis­ten­tial risks to so­ciety” if de­plo­yed without ade­qua­te sa­fe­guards. That's not to say Clau­de is cu­rrently black­mai­ling cus­to­mers. Anth­ro­pic has ob­ser­ved some of this beha­vior in tes­ting, but it also iden­ti­fies more im­me­dia­te se­cu­rity con­cerns as it gives its mo­dels ac­cess to ex­ter­nal tools, code re­po­si­to­ries, and cor­po­ra­te sys­tems. Anth­ro­pic warns that at­tac­kers could hide ma­li­cious ins­truc­tions in web­si­tes, do­cu­ments, or API res­pon­ses that its AI agents can ac­cess. If an agent fo­llows those ins­truc­tions, it could use a cus­to­mer's per­mis­sions to steal data, tam­per with code, or make unautho­ri­zed transac­tions. Even tes­ting whether the mo­dels will beha­ve them­sel­ves is pro­ving tricky. Anth­ro­pic rec­kons some mo­dels beha­ve dif­fe­rently when they de­tect or sus­pect they're being eva­lua­ted, ma­king it har­der to pre­dict what they will do outsi­de the tes­ting en­vi­ron­ment. Kee­ping its mo­dels on a short leash could also cost Anth­ro­pic bu­si­ness. The com­pany ack­now­led­ges that hol­ding back more ca­pa­ble sys­tems could give com­pe­ti­tors an ad­van­ta­ge, so­mething it has al­ready ris­ked by res­tric­ting ac­cess to its Mythos model th­rough Pro­ject Glass­wing be­cau­se of its abi­lity to un­co­ver se­cu­rity vul­ne­ra­bi­li­ties. The foun­ders aren't pro­mi­sing to put sha­rehol­ders first, either.