Teridentifikasi, Ini Identitas 4 Korban Tewas Kecelakaan Panther Tertabrak Kereta Api di Prabumulih

Pilihan motor listrik di bawah Rp20 juta semakin beragam dengan spesifikasi dan fitur untuk mobilitas harian.

227 install commands were found in corporate docs pointing at code nobody owns.

Next.js patches two critical unauthenticated RCE flaws affecting Windows deployments and sites with AVIF optimization enabled.

Polda Sumut bongkar motif mengerikan pembunuhan driver online Riyan Alamsyah di Medan. Pelaku nekat menjerat korban demi uang beli sabu. Terancam hukuman mati!

Israel mencopot dan mengubah papan nama Gerbang Damaskus menjadi Gerbang Kuno. Berikut berita lengkapnya.

Larasati Aulia, alumni SMA Unggulan CTARSA Foundation Sukoharjo, berhasil meraih BSI Scholarship Unggulan untuk kuliah di ITB.

For much of the last year, Alphabet Inc. has been the Big Tech stock to beat as investors bet that it was the most-likely winner from the artificial intelligence boom.

Australian police charged two men over their alleged TeamPCP role in supply chain compromises affecting Trivy, KICS, and LiteLLM.

Perusahaan investasi Amerika Serikat, Morgan Stanley, kembali memborong saham GOTO.

Pasar properti residensial RI melambat pada Kuartal II-2026 dengan skor 50,3. Pinhome mencatat perubahan perilaku calon pembeli.

Dua karyawan perusahaan pengolahan uang ditangkap karena menilap uang dari ATM bank pelat merah. Kerugian mencapai Rp10,15 juta.

Film The Last House telah hadir di Netflix sejak 7 Agustus 2026. Mengusung perpaduan horor, thriller, dan fiksi ilmiah, film garapan Louis Leterrier ini bercerita tentang sebuah keluarga yang menghadapi situasi tidak biasa ketika mereka mendapati diri mereka tidak mampu meninggalkan rumah.

Menkes Budi Gunadi Sadikin menjelaskan pemulihan layanan kesehatan di NTT pasca-bencana. Fokus pada identifikasi pasien dan perbaikan fasilitas kesehatan.

Spark RAT targets Cambodia through a multi-stage campaign that uses a vulnerable OPSWAT driver to terminate security processes.

Ketua Umum DPP Partai Ummat Ridho Rahmadi mengundurkan diri dari jabatannya. Jabatan Plt Ketua Umum DPP Partai Ummat kini diemban Ustaz Ansufri Idrus Sambo. Ketua...

Hingga kini, identitas maupun spesifikasi varian baru BMW M2 itu belum diketahui. Namun, informasi mengenai rencana tersebut mulai mencuat

Persita Tangerang mengusung target finis di papan atas klasemen akhir Super League 2026-2027.

AS memasukkan organisasi non-profit yang berbasis di Inggris, Palestine Action, sebagai kelompok teroris.

Pembangunan tanggul laut di pesisir Pantura Jawa Tengah dipercepat guna mengatasi banjir rob yang diprediksi BMKG berlangsung hingga akhir Agustus

Kemenkeu mengklarifikasi pernyataan Purbaya soal anggaran sebesar Rp7 triliun BPS. Dana tersebut bukan dikucurkan hanya untuk membiayai perbaikan DTSEN.

Kakorlantas Polri Irjen Wibowo menggelar Gebyar Keselamatan Lalu Lintas untuk meningkatkan keterampilan pengemudi ojol, termasuk pelatihan PPGD.

Transformasi industri properti Indonesia semakin mengarah pada pembangunan yang tidak hanya mengutamakan desain dan kenyamanan, tetapi juga efisiensi energi ...

Simak jadwal lengkap Persib Bandung di Super League 2026-2027.

Dua bintang Korea, Son Ye-jin dan Ji Chang-wook, akan berperan dalam drama Netflix "The Scandal".

Audio-Technica ATH-M20xSTS resmi diperkenalkan sebagai headphone untuk streamer, gamer, dan kreator konten. Berbasis ATH-M20 dengan mic kondensor.

Perdana Menteri Israel Benjamin Netanyahu mengatakan dalam pembicaraan tertutup bahwa partai Likud-nya akan kalah dalam pemilihan Knesset (parlemen) mendatang kecuali...

Kualitas udara di 4 wilayah Kalsel masih kategori sedang. Dokter mengingatkan penderita asma dan kelompok rentan mewaspadai paparan kabut asap.

peserta mendapatkan edukasi mengenai manfaat program BPJS Ketenagakerjaan, mulai dari perlindungan terhadap risiko kecelakaan kerja hingga m

PT Bank Tabungan Negara (Persero) Tbk. (BBRI) menargetkan penyaluran KPR Rp200 miliar pada Danantara Housing Expo 2026 dengan penawaran bunga mulai 2,75%. Stabilitas.id - PT Bank Tabungan Negara (Persero) Tbk.…

Temukan 5 jenis teh yang efektif untuk membantu penurunan berat badan. Dari teh hijau hingga pu-erh, masing-masing memiliki manfaat unik untuk diet sehat.

GPUThor defeats ECC on NVIDIA GDDR6 workstation GPUs and demonstrates host privilege escalation to root on an RTX A6000.

Bencana banjir bandang dahsyat dilaporkan melanda wilayah perbatasan Nepal-Tibet, China. Peristiwa mematikan ini menyapu kota, jalan, hingga jembatan vital.

Kementerian ESDM dan Pertamina memprediksi konsumsi BBM bersubsidi dan LPG 3 kg bakal melebihi batas atau jebol hingga akhir 2026.

Generasi terbaru BMW X7 G67 tertangkap kamera sedang menjalani pengujian di Sirkuit Nürburgring, Jerman

Data BPS menunjukkan jutaan rumah tangga masih membutuhkan hunian layak.

Minnesota Timberwolves menyetujui kontrak dua tahun senilai 13 juta Dolar AS untuk mendapatkan Jonathan Kuminga pada hari Rabu (26/8) waktu AS, menurut Shams Charania dari ESPN. Langkah ini dilakukan setelah Hawks secara resmi menolak opsi tim Kuminga pada bulan Juni, menjadikannya pemain bebas setelah kepergiannya yang penuh gejolak dari Golden State Warriors musim lalu. Kuminga dilaporkan memilih Timberwolves daripada tawaran serupa dari Los Angeles Lakers, Chicago Bulls, dan Portland Trail Bl

Respons Resmi Manajemen Persija Jakarta Terkait Polemik Kenaikan Harga Tiket Peluncuran Tim

Tim penyelamat bergegas untuk menemukan lebih dari 1.300 orang yang hilang di perbatasan Nepal-Tibet akibat banjir bandang dan tanah longsor.

Departemen Kehakiman AS menghidupkan kembali sistem pengadilan maritim kuno yang telah lama tidak digunakan untuk mempercepat penyitaan kapal tanker minyak Iran...

Sedimentasi parah di Sungai Cacaban memicu kekhawatiran banjir warga Tegal. DPRD didesak kawal normalisasi sungai dan perbaikan infrastruktur terkait.

Aturan penggunaan media sosial bagi anak semakin ketat di berbagai negara. Simak daftar negara dan batas usia yang diterapkan.​

Penerbitan SEMA No 3 Tahun 2026 tentang pelimpahan dan penyelenggaraan pemeriksaan pokok perkara dalam hal terdapat praperadilan merupakan langkah yang tepat. Penerbitan...

Israel mengusir perwakilan Belanda dari Pusat Dukungan Internasional untuk Gaza gegara Amsterdam melarang impor dari permukiman ilegal di wilayah Palestina.

DPR RI dijadwalkan tetap menggelar rapat paripurna di tengah pelaksanaan massa aksi yang dilakukan sejumlah organisasi masyarakat di depan kompleks parlemen.

Reaksi Lucu Presiden FIFA Gianni Infantino Saksikan Drama Gol Bunuh Diri di Final Piala AFF 2026

Dolly Parton meninggal pada 25 Agustus 2026, namun warisannya terus hidup.

Presiden China Xi Jinping mengerahkan pejabat dan tim penyelamat ke Kabupaten Gyirong, Kota Shigatse, Tibet, untuk menangani banjir bandang dan longsor lumpur yang menewaskan tiga ANTARA News kl 18 ...

Hasil uji kelayakan pimpinan BI, inflasi PCE AS, dan pertumbuhan ekonomi AS akan menjadi penggerak pasar hari ini.

Wakil Menteri Nguyen Van Long meminta agar informasi catatan kriminal yang ditampilkan di VNeID dioperasikan secara cerdas, diperbarui secara otomatis dalam waktu nyata, dan selalu dalam keadaan sah secara hukum untuk memastikan warga dapat mengaksesnya segera setelah membuka aplikasi; sama sekali tidak boleh ada situasi di mana orang harus berulang kali mengajukan permohonan penerbitan ulang catatan kriminal; dan pada saat yang sama, hal itu harus melayani hak-hak sah masyarakat secara maksimal.

Timnas Indonesia siap berjaya di FIFA ASEAN Cup 2026 yang berlangsung di Jakarta dan Bandung pada 21 September hingga 4 Oktober 2026.

Nvidia reported results that beat expectations, but it also expects revenue to grow another 70% in fiscal 2028.

Banjir bandang di perbatasan Nepal-Tibet menewaskan 157 orang, dengan banyak korban hilang. Upaya penyelamatan terkendala, PBB dan India siapkan bantuan.

Banjir bandang yang melanda wilayah Nepal sejak Rabu (26/8) menewaskan lebih dari 150 orang dan 384 lainnya hilang.

Banjir bandang dahsyat menerjang kawasan perbatasan Nepal dan China pada Rabu (26/8/2026). Sedikitnya 160 orang tewas dalam bencana yang terjadi secara tiba-tiba tersebut, sementara ratusan lainnya…

Kabar duka berpulangnya ibunda Deddy Corbuzier, Heniwaty, menyisakan cerita perjuangan medis yang luar biasa.

Berlibur ke Eropa lebih mudah bagi pemegang paspor Indonesia. Temukan 10 negara yang menawarkan bebas visa, eVisa, dan VoA untuk perjalanan Anda.

Beragamnya pilihan mobil keluarga di pasaran saat ini bisa membuat kamu merasa kebingungan.

Earlier this year, the bears thought advances in AI models could disintermediate the cybersecurity sector. So far, they've been proven wrong.

The platform updated some parts of the format while preserving the functionality of this previously retired CTA feature.

AdAge reported that the app is teaming up with creators to highlight the benefits of shared viewing experiences, such as watching Reels with friends.

Bencana longsor di perbatasan Nepal-Tibet menyebabkan banjir bandang, menewaskan 157 orang dan ratusan wisatawan hilang. Upaya penyelamatan terus dilakukan.

Enam kelompok masyarakat menerima persetujuan Perhutanan Sosial. Pemerintah meminta kawasan kelola dijaga dengan baik jangan sampai terjadi kebakaran hutan dan lahan.

Banjir bandang di Nepal telah menewaskan 157 orang dan ratusan lainnya hilang. Seluruh pemukiman hancur, sementara pencarian terus dilakukan.

Periset meninggalkan pekerjaan bergaji Rp 2,5 miliar di Huawei dan pulang kampung membantu keluarganya bertani jagung.

Film Operasi Pesta Copet tayang 27 Agustus 2026. Bercerita sekelompok pemuda mencopet di festival musik, film ini menggabungkan aksi, komedi, dan kritik sosial.

Ditlantas Polda Metro Jaya menyiapkan rekayasa lalu lintas di sekitar Gedung DPR/MPR guna mengantisipasi kemacetan imbas aksi 27 Agustus pada Kamis (27/8).

Kuasa hukum Febrie Adriansyah menyoroti pentingnya independensi hakim dalam putusan praperadilan. Simak selengkapnya di sini.

Lokasi gunung emas itu berada di Provinsi Kivu Selatan, Venant Burume Muhigirwa, pernah menjadi buruan para penambang untuk mencari kekayaan.

Penurunan berat badan melalui diet dapat menstabilkan tekanan darah pada hipertensi. Dokter menjelaskan manfaatnya bagi kesehatan jantung dan peradangan.

Anthropic will rent around 460 megawatts of compute capacity at an Nscale data center development in West Virgina, according to sources.

In extensive interviews, the leaders of the company that ushered in the AI boom lay out their vision for its future

FBI disrupted proxy infrastructure used in China-linked espionage to profile and steal data from U.S. critical infrastructure and other sectors

PT TASPEN menerima kunjungan delegasi Timor-Leste untuk berbagi pengalaman dalam pengelolaan jaminan sosial dan sistem pensiun ASN.

GAC menawarkan Special Price mobil listrik di GIIAS Surabaya 2026, dengan diskon hingga Rp65 juta untuk AION V Luxury.

Peneliti Matthijs Wibier merekonstruksi fragmen buku hukum Romawi abad keempat yang disimpan di Penn Museum.

Pengelolaan nilai manfaat dana haji bukan sekadar persoalan teknis-finansial.

I.League akan mengupayakan agar League Cup menjadi jalan klub Indonesia tampil di kompetisi Asia.

The most critical bottleneck isn't power or turbines, Melius Research says, but regulation.

Iran menyerukan kepada pemerintah berbagai negara untuk menolak dan mengecam penerapan ekstrateritorial dari tindakan koersif sepihak AS. Teheran juga memperingatkan...

Taksi listrik Green SM terlibat kecelakaan dengan kereta di Bekasi Timur dan Jakarta pada tahun 2026 ini.

PENGABDIAN: Poltekkes Kemenkes Manado telah melaksanakan kegiatan Pengabdian kepada Masyarakat bertemakan "Pengembangan Wirausaha Sosial di

TPO - Pada tanggal 26 Agustus, di Kota Da Nang, Surat Kabar Konstruksi menyelenggarakan Forum "Pariwisata dan Resor Real Estat Vietnam 2026". Forum ini mempertemukan banyak lembaga manajemen, pakar, bisnis, dan investor untuk menilai situasi saat ini, mengidentifikasi peluang, dan menemukan solusi pembangunan berkelanjutan untuk pasar.

Partai Ummat menggelar diskusi yang bertujuan sebagai ruang pertemuan antara gagasan hukum dan ketatanegaraan dengan nilai-nilai moral.

Kuasa hukum eks Jampidsus Febrie Adriansyah menilai proses hukum kliennya cederai keadilan prosedural dan langgar *due process of law*. Putusan dibaca Jumat.

Dokter spesialis saraf menjelaskan bahwa penyakit Parkinson tidak hanya menyerang lansia tetapi juga usia muda akibat faktor genetik dan lingkungan

Mantan pemain timnas Inggris Raheem Sterling dilaporkan terancam hukuman penjara setelah didakwa atas tiga pelanggaran terkait kecelakaan lalu lintas di ...

A wave of cyberattacks against critical infrastructure raises questions about their preparedness.

Deretan SUV tersebut hadir dengan beragam teknologi. Mulai dari mesin pembakaran internal (ICE), Plug-in Hybrid Electric Vehicle (PHEV), hingga full EV

Kementerian ATR/BPN temukan 11 juta rumah MBR belum bersertifikat. Pemerintah berikan sertifikat gratis dengan syarat tertentu bagi MBR.

“Siapa yang bisa kamu percaya?” Pertanyaan tersebut menjadi pembuka yang menggambarkan nuansa penuh kecurigaan dalam film horor terbaru “Hasut”. Film ini menawarkan teror yang tidak hanya datang dari suasana mencekam, tetapi juga dari rasa ragu terhadap orang-orang di sekitar.

Motor City tayang mulai 26 Agustus 2026. Cinta, pengkhianatan, dan balas dendam era 70-an bikin penonton tegang hingga bertepuk tangan.

Dolly Parton meninggal di usia 80 tahun. Pesan terakhirnya, "Aku akan selalu mencintaimu," menggambarkan warisan dan pengaruhnya yang abadi dalam musik.

Presiden AS Donald Trump memerintahkan pengibaran bendera setengah tiang selama satu pekan usai musisi Dolly Parton tutup usia pada Selasa (25/8).

Pemerintah dan DPR perlu menempatkan RUU HPI sebagai bagian dari agenda yang lebih besar yaitu pembaruan sistem hukum perdata nasional secara menyeluruh. RUU HP

Banjir bandang dahsyat melanda wilayah pegunungan Nepal yang berbatasan dengan Tibet pada Rabu (26/8/2026).

Pemerintah Kabupaten (Pemkab) Parigi Moutong (Parimo), Sulawesi Tengah menyalurkan makanan siap saji kepada warga terdampak banjir bandang di Desa Avolua, Kecamatan Parigi ANTARA News sulteng 1 ...

Dengue membebani ekonomi Indonesia hingga Rp9 triliun. ADINKES 2026 dorong kolaborasi lintas sektor, integrasi data, dan pencegahan terpadu perkuat layanan primer.

Rumor soal penjualan saham tambang batu bara PT Bayan Resources Tbk (BYAN) kembali mencuat.

Para penggugat menuding perusahaan media sosial turut berkontribusi terhadap krisis kesehatan mental di kalangan anak muda di Amerika Serikat

Polrestabes Semarang menangkap dua pria yang menyamar sebagai wanita di Tinder untuk memeras korban.

Timnas Bangladesh meminta FIFA menunda laga penentuan posisi di FIFA ASEAN Cup 2026.

In an hourlong interview, he said his industry was downplaying the risks, which he believes include mass unemployment and bioterrorism.

Xiaomi akhirnya resmi mengumumkan SoC terbaru buatan mereka sendiri yang akan dipakai di smartphone mereka yaitu Xring o3. SoC ini akan dibenamkan di

Yogi Gilang, 39, ditangkap karena menjual layanan internet Starlink tanpa izin di Mentawai. Kasusnya kini di pengadilan, dengan kejanggalan prosedur hukum.

Menteri PU Dody Hanggodo melaporkan penanganan jalan, jembatan, air bersih, dan fasilitas pemerintah terdampak gempa NTT kepada Presiden Prabowo.

Di kesempatan tahun 2026 ini, Zona Serayu Utara menjadi target utama untuk aktivitas mapping mandiri untuk Mahasiswa Teknik Geologi UGM angkatan tahun 2024 di akhir semester 4 ini. Kali ini menyusuri daerah yang cukup bisa dibilang lumayan jauh dari aktivitas magmatisme yakni di zona Kendal. Secara umum, daerah Kendal dipenuhi dengan lapisan sedimen yang terlipat […]

Mengemudi dalam kondisi mabuk disebut sama saja dengan berhalusinasi oleh pakar keselamatan berkendara.

Saham emiten pertambangan nikel PT Abadi Nusantara Hijau Investama Tbk (PACK) melesat dalam beberapa pekan terakhir.

RUPSLB PT Harta Djaya Karya Tbk (MEJA) pada 21 Agustus 2026 menyetujui peningkatan modal dasar dari semula sebesar Rp100 miliar menjadi Rp208,65 miliar.

Meski sudah berakhir, namun kamp mini yang dibuat Luka Doncic di Slovenia masih menjadi pembahasan menarik. Salah satunya ketika muncul sosok legendaris dalam latihan tim Los Angeles Lakers tersebut. Sosok tersebut adalah Holger Geschwindner.

Indonesia menempati peringkat ke-12 di Asia dalam kemampuan bahasa Inggris dengan skor 471, tertinggal dari Malaysia dan Filipina.

Pemerintah Provinsi Sulawesi Tenggara (Sultra) menyiapkan pembebasan lahan untuk mendukung rencana pembangunan tanggul pengendali banjir di Sungai Wanggu dan Sungai Lepo-Lepo, Kota Kendari.

Pemerintah akan segera mengoreksi dan memperbaiki data anomali desil setelah mendapat banyak keluhan masyarakat, serta akan berdiskusi dengan BPS dan DEN.

Seorang profesor di Universitas Islam Internasional Malaysia (IIUM) dipecat usai melontarkan klaim viral soal orang Melayu kuno bisa terbang.

Seorang pria bernama Andre tewas setelah berkelahi dengan Abu Sidik di kebun sawit. Perkelahian dipicu oleh dugaan pencurian buah sawit.

CISA adds CVE-2026-60004 to KEV amid active Gitea RCE exploitation; a separate reported attack deployed a miner-like dropper.

Israel ketar-ketir usai Turki mengumumkan memiliki rudal balistik hipersonik baru yang diperkirakan dapat menjangkau seluruh wilayah Negeri Zionis.

PSI menilai wacana pergantian pemerintahan di tengah masa jabatan yang sempat dilontarkan Budi Arie sangat bertentangan dengan prinsip partai.

Menguji Kedudukan Advokat dan Pihak Lain dalam PMK 44 Tahun 2026 dari Perspektif Legalitas, Kepastian Hukum, dan Perlindungan Wajib Pajak

Kecelakaan maut melibatkan dua sepeda motor terjadi di Jalan Kemusuk-Sidomulyo, tepatnya di depan Gapura Dusun Sembuh Kidul, Sidomulyo, Godean, Sleman, Selasa (25/8) petang. Dalam kejadian tersebut, satu orang meninggal dunia dan seorang lainnya mengalami luka-luka.

Anggota Komisi VIII DPR RI Selly Andriany Gantina meminta Kementerian Sosial (Kemensos) memastikan pemberian perlindungan sosial bagi korban gempa bumi di Nusa Tenggara Timur ANTARA News ntt 2 ...

Timnas Vietnam bisa merayakan gelar juara Piala AFF 2026 di hadapan Presiden FIFA Gianni Infantino.

Kemenkes siapkan aturan untuk lindungi anak sekolah dari kualitas udara buruk akibat karhutla. Pembelajaran tatap muka dihentikan saat ISPU tidak sehat.

Serangan rudal-rudal dan drone-drone Iran diklaim membuat pusat data intelijen Amerika Serikat di sejumlah negara Timur Tengah porak-poranda.

Para pejabat militer AS memperingatkan operasi tempur berskala besar dapat menguras pasokan amunisi dan rudal pencegat Patriot secara mengkhawatirkan.

Dewan Pimpinan Pusat Partai Solidaritas Indonesia DPP PSI membantah keterlibatan Budi Arie Setiadi di dalam struktur partai

Partai Buruh tidak akan menggelar aksi besar pada Agustus–September 2026 sambil mengawal revisi outsourcing, JHT, pesangon, dan UU Ketenagakerjaan.

Yogi Gilang Arya Setia didakwa menjual kembali akses internet Starlink di Mentawai. Kuasa hukum menilai harusnya diselesaikan secara administrasi.

Cewek mabuk pengemudi Outlander, menyesal usai menewaskan warga Surabaya di kecelakaan akibat alkohol. Dia kini jadi tersangka dengan ancaman 6 tahun penjara.

Harga emas nyaris tak bergerak setelah sempat menembus level tertinggi dalam lebih dari tiga bulan.

Seorang perempuan di China alami kesulitan berjalan setelah 15 hari hanya minum air putih. Tren 'bigu' di China jadi salah satu pemicunya.

Antrean panjang mengular terpantau di sejumlah pom bensin di Teheran, ibu kota Iran, pada Selasa (25/8), menyusul langkah AS mengumumkan rentetan sanksi baru.

Vietnam menghadapi leg kedua final Piala AFF 2026 dengan modal positif berkat kemenangan 2-0 di kandang Thailand.

Badan Penanggulangan Bencana Daerah (BPBD) melaporkan sekitar 12 unit rumah warga di Desa Avolua, Kabupaten Parigi Moutong, Sulawesi Tengah rusak akibat ...

Universitas di Malaysia pecat profesor buntut sederet pernyataan kontroversial soal sejarah Melayu.

Berikut dua pemain yang didatangkan Manchester United setelah Carlos Baleba, ada Ivan Toney dan Lewis Hall.

Komponen mobil seperti sensor, alternator, ECU, throttle body, dan koil rentan rusak jika terkena air serta mengganggu kelistrikan dan mesin.

Pemblokiran rekening Bank Mandiri milik Supriyono disorot YLKI dan didalami OJK, menyusul permintaan aparat penegak hukum kepada bank.

Empire State Building menyala dengan kilauan warna merah muda untuk menghormati Dolly Parton.

PSI mengkritik gaya politik Budi Arie yang, menurutnya, mengedepankan manuver dan spekulasi demi kepentingan jabatan.

Komite Rakyat Provinsi baru saja mengeluarkan surat edaran resmi yang berfokus pada upaya mengatasi dampak Topan No. 4, hujan lebat, banjir, tanah longsor, dan secara proaktif mencegah serta memerangi bencana alam di masa mendatang (Surat Edaran Resmi No. 03/CĐ-UBND tanggal 25 Agustus 2026).

Provinsi Dong Thap berfokus pada penerapan solusi mendesak dan jangka panjang untuk mengatasi situasi erosi tepi sungai, kanal, dan parit yang semakin kompleks. Terutama, erosi di daerah pedalaman telah meningkat tajam, menyebabkan kerusakan pada rumah, harta benda, produksi, dan infrastruktur.

Dinas Kesehatan (Dinkes) Kabupaten Manggarai Barat, Nusa Tenggara Timur (NTT), mencatat 3.151 kasus penyakit di wilayah terdampak gempa.

Ledakan AI membuat kebutuhan konektivitas melonjak. Telkom mengungkap kapasitas kabel laut Bifrost habis kurang dari setahun setelah dibangun lima tahun

KORLANTAS POLRI, Pesawaran - Satuan Lalu Lintas (Satlantas) Polres Pesawaran Polda Lampung gerak cepat tangani kecelakaan lalu lintas yang melibatkan sepeda ...

Berat badan susah turun padahal sudah diet dan rajin olahraga bisa membuat seseorang merasa frustrasi.

Manchester United resmi mendatangkan Carlos Baleba dari Brighton & Hove Albion pada bursa transfer musim panas 2026.

Partai Hanura ganti lambang singa jelang Pemilu 2029. Apakah perubahan logo ini efektif mendongkrak elektabilitas atau hanya menarik perhatian?

Jika otoritas China menentukan seorang warga Taiwan di China bekerja untuk perusahaan yang dianggap musuh, orang tersebut terancam tak bisa pulang. Otoritas keamanan...

Kecelakaan lalu lintas terjadi di Jalan Gatot Subroto, Jakarta Pusat, melibatkan mobil boks dan sepeda motor. Petugas sedang menangani insiden ini.

LIGHThouse Clinic meluncurkan dua pilihan terapi terbaru bernama TirzePen & TirzeSlim berbasis senyawa tirzepatide guna membantu pasien mengendalikan f...

A similar Reply to Keywords feature is already available via inbox automation and Meta Business Agent options.

BoP desak Hamas setop terbangkan layangan ke wilayah Israel, juga minta Israel patuhi gencatan senjata.

Dua bulan jelang pemilihan, partai pimpinan Perdana Menteri Israel Benjamin Netanyahu meraih perolehan jeblok dalam survei terbaru.

HEADLINEKALTIM.CO, TENGGARONG– Jalan poros Dusun Oloy Kecamatan Muara Muntai yang menghubungkan antar desa dan  kecamatan menuju Kota Tenggarong yang alami longsor parah sepanjang 10 meter akan ditangani Pemkab Kukar melalui Dinas Pekerjaan Umum (PU). “Karena itu akses penting, makanya akan kita tangani, jalan tersebut merupakan jalur utama antar desa dan kecamatan di Kukar,” ucap Kabid Bina Marga, Dinas PU Kukar,..

Sedikitnya 53.971 rumah rusak akibat gempa di NTT. Warga Nagekeo masih memilih tidur di tenda karena khawatir gempa susulan.

Penulis: Muhammad Anshori

Data desil yang tak akurat menghambat akses bantuan sosial bagi yang berhak. Mahasiswa terancam putus kuliah dan keluarga miskin terdaftar kelas atas.

The AI startup is likely to top SpaceX’s eye-popping potential revenue estimate.

Sebanyak 20 pelaut dan pekerja pelabuhan tewas dalam 68 insiden di kawasan Teluk dan perairan sekitar Selat Hormuz selama enam bulan perang antara Iran dan AS.

This is our second sale in as many days to reduce our exposure to the theme.

Performa mesin merupakan salah satu faktor utama yang dapat menentukan kenyamanan dan keamanan dalam proses berkendara.

By Mike Scarcella WASHINGTON, Aug 25 (Reuters) - Alphabet's Google on Tuesday expanded its Gemini Enterprise artificial intelligence platform with new tools for lawyers and law firms, accelerating a

Pria yang wajahnya disebut mirip Pangeran Harry kebanjiran tawaran pekerjaan setelah Harry dan Meghan Markle dikabarkan akan kembali tinggal di Inggris.

Kanwil Kemenkum Babel Perkuat Pembinaan Posbankum dan Penyelenggaraan Bantuan Hukum Triwulan III Tahun 2026

Kenali penyebab demam reumatik pada anak yang dipicu bakteri Streptococcus Grup A Simak gejala risiko penyakit jantung reumatik dan cara pencegahannya

Namun, pengalaman Mulyani menunjukkan bahwa pemeriksaan calon pembeli tidak berhenti sampai di situ.

Sinopsis Agito: Psychic War, film sci-fi terbaru Kamen Rider yang menghadirkan kembali Makoto Hikawa dan Shoichi Tsugami di bioskop Indonesia.

OpenAI has spent the past couple years quietly building “Jalapeño,” an inference chip just announced at Hot Chips. Rumors of a successful tapeout had been swirling for a while. But now we have details. OpenAI invited us to look at their chip, go to their labs to check out how real it is, and benchmark it with our InferenceX suite.

In June, OpenAI unveiled the chip program in partnership with Broadcom, built from a blank slate exclusively for LLM inference. Design work began in the middle of 2024, going from initial team hiring to manufacturing tape-out in ~16 months, an extremely fast ASIC development cycle.

In general first generation chips are not competitive, but OpenAI bucks the trend by being industry leading and beating every Nvidia, AMD, and Google chip we have been able to test on multiple top open source models. OpenAI does this with extreme hardware software codesign. Surprisingly, OpenAI is not over specialization on any specific part of model inference, but instead by focusing on being a general chip that delivers high performance in all scenarios.

In this article, we will go into architectural details, software details and performance results for Jalapeño on InferenceX.

Everyone says that OpenAI’s chip is specialized for OpenAI models, but that’s wrong, OpenAI made a generalized chip for AI inference.

The timelines are insane. It shows that claims that use of AI is being used to accelerate chip design are real. Regardless of the quick timelines,Open AI spent a bunch of money, made pragmatic design decisions and their team is cracked, so this comes as no surprise.

Just looking at the specs, it is an immediate contender:

And the use of HBM4 makes it stand out as comparable to flagship GPUs from NVIDIA and AMD:

A lot of the media coverage of this chip has followed a few throwaway comments from OpenAI that claim the chip will be optimized for their models in a way that other chips are not. This is wrong. Jalapeño is a generalized inference chip capable of running all sorts of models, and all sorts of workloads, including our benchmark InferenceX, where we ran the benchmark with OpenAI engineers in the lab. As a joke, OpenAI even showed us it running Doom, which was ported to their chip with just Codex prompts.

The following is our headline perf/W result, looking at token throughput per All-in utility MW. Jalapeño smokes every other chip. All this is done without Multi Token Prediction (MTP), while the other chips on the chart are the best performing configs of each respective SKU, all with MTP.

Jalapeño beats Blackwell on perf/W across almost all scenarios without being tuned for any specific point in the curve. It excels not only in low-latency scenarios but also in high-throughput scenarios. A more apples to apples comparison is against Single Token Prediction results, it knocks every competitor out of the water. At low concurrency scenarios, Jalapeño demonstrates remarkable interactivity, hitting over 700 tokens per sec per user at concurrency 1 on the DeepSeek R1 model.

Incredibly, this is all achieved with single-token prediction (STP), no speculative decoding and no prefill-decode disaggregation. In addition to DeepSeek R1, we also got to see some other models, including Kimi-K2.5 and GPT-OSS which ran at approximately 1,400 tok/sec/user. For all models, we confirmed that Jalapeño’s GSM8k evals attained results on par with Nvidia chips.

Some caveats on this. First, all numbers are provided to us by OpenAI. We verified the InferenceX runs in person in the lab, but we did not run the full suite of InferenceX benchmarks nor have we seen AgentX results. AgentX is our preferred suite for comparing chip performance due to the datasets’ long context and multi-turn characteristics that reflect the cache behavior of realistic production workflows. Frameworks that perform well on 8k1k may perform worse on AgentX as real production loads stress components like routers, prefix cache mechanisms, cache management, offload infrastructure, etc. These are not tested by single turn 8k1k. Read more about this in out AgentX article.

[

](https://newsletter.semianalysis.com/p/agentx-inferencexv3-does-cuda-moat)

[

AgentX - InferenceXv3: Does CUDA Moat Hold up in Agentic Inferencing?

](https://newsletter.semianalysis.com/p/agentx-inferencexv3-does-cuda-moat)

Cam Quilici, Bryan Shan, and 5 others

·

Aug 24

[

Read full story

](https://newsletter.semianalysis.com/p/agentx-inferencexv3-does-cuda-moat)

Second, we believe that comparison to Blackwell is somewhat incomplete and unfair. Jalapeño is really competing against chips like Rubin that also use HBM4. Vera Rubin systems are starting to ship to customers right now, while it will still be some time before OpenAI has anything beyond engineering samples of Jalapeño.

Thus, performance should really be compared against Rubin, not Blackwell, and in some sense we expect a custom chip like Jalapeño to outperform Blackwell. Vera Rubin NVL72 delivers 5.4x the perf/MW of GB200 NVL72 as we described in our article analyzing the NVIDIA performance claims in their launch with CoreWeave last month. We will compare Jalapeño to Vera Rubin’s July performance figures later below.

[

](https://newsletter.semianalysis.com/p/vera-rubin-nvl72-vs-gb200-nvl72-inference)

[

Vera Rubin NVL72 vs GB200 NVL72? Inference TCO & Architecture Analysis

](https://newsletter.semianalysis.com/p/vera-rubin-nvl72-vs-gb200-nvl72-inference)

Alec Ibarra, Bryan Shan, and 6 others

·

Jul 23

[

Read full story

](https://newsletter.semianalysis.com/p/vera-rubin-nvl72-vs-gb200-nvl72-inference)

Third, the models being tested are not on the open frontier. NVIDIA and AMD have published results on larger models such as DeepSeek V4 Pro and Kimi K3, using AgentX. The larger the model and the more recent the release, the more complicated it is to bring up on a new chip. With that said the models OpenAI has working on Jalapeno aren’t exactly small either.

OpenAI designs for perf/W. The reason is simple: OpenAI is currently limited by datacenter power, not by budget or floorspace, and thus tokens per MW is paramount. At Computex 2026, Jensen said that perf/W, reliability and long lifetime are the core features of future GPUs. To quote: “If you have 1 gigawatt of power, then throughput per watt is revenue”. He also mentioned that choosing the wrong architecture just because the chips are cheaper doesn’t make sense.

This was emphasized by Nvidia during the Vera talk at Hot Chips 2026 while showing the same revenue graph: “The data center is power limited today.” Power matters and drives revenue.

Operators cannot simply obtain more MW because adding GPUs and adding grid capacity happen on very different timescales. Datacenter power envelopes have constraints such as their utility interconnection, infrastructure, cooling capacity, and UPS/backup-generation design. Grid delays repeatedly outpace hardware and construction timelines, driving the need for BtM (behind-the-meter) power capacity: gas turbines and on-site generators built and located at the data center itself. This capacity sits behind the utility’s meter rather than being drawn from the public grid. It lets an operator power a facility without waiting on grid interconnection and utility upgrades, which is exactly why xAI’s Colossus 2 relies so heavily on BtM while its actual grid connection lags far behind. Find out more in our Energy model.

As we wrote in an X post, tok/s/MW reduces to tokens per joule since a watt is a joule per second. This makes tok/s/MW representative of a system’s efficiency and ability to convert energy into tokens.

On this front, even when compared with Rubin, Jalapeño wins. OpenAI’s Jalapeño has STP output token throughput per MW surpassing Vera Rubin’s MTP results that NVIDIA and CoreWeave published in July. It also far exceeds GB200’s 2025 MTP results. As mentioned in our Vera Rubin article, VR was compared to 2025 GB200 results because that was a similar stage of early bring-up, and comparing to GB200 in 2025 holds software maturity constant. Following this logic, we compare Vera Rubin’s latest July 2026 results, GB200 2025 results, and today’s Jalapeño results. This is a very valid comparison as these are the best public Rubin numbers, and OpenAI taped out their chip after Rubin. Both OpenAI and Rubin are still immature thus performance will continue to rise.

On perf/TCO, Vera Rubin and Jalapeño are head-to-head, producing almost the same number of output tokens per $. However, as previously mentioned, Jalapeño’s results are obtained without speculative decoding and Vera Rubin’s results use speculative decoding. Speculative decoding adds a 3x+ improvement is cost per token. When speculative decoding is implemented on Jalapeño, this will mean Jalapeño will serve much lower cost tokens. Of course, part of this TCO advantage comes from trading Nvidia’s high margins for Broadcom’s lower (though still high) margins. But this is not all of it. For example, Meta and Microsoft’s AI ASIC programs not getting off the ground despite being at it for much longer shows that cost is only one part of the equation. For Jalapeño’s full TCO breakdown, see the SemiAnalysis AI Cloud TCO model.

Architecturally, OpenAI chose not to disaggregate prefill and decode (PD) across separate chip pools. The draft model and main model share the same chips and fabric, a design philosophy that trades some theoretical efficiency for practical operations. The motivation is that the workload mix changes over time, for example the ratio of input to cache write to cache read to output tokens has changed significantly as we have moved through the three eras of models (knowledge, reasoning, and agentic, as discussed in our recent article). Therefore, picking a fixed amount of heterogenous prefill silicon and decode silicon up front can lead to inefficiencies over time. OpenAI chooses a homogenous pool in this architecture and tries to make the chip perform well on everything.

And it does. On Kimi K2.5 (which Cursor Composer 2.5 is based on), Jalapeño reaches nearly 700tok/s/user and more than 9x the next best performing chip at 100tok/s/user.

On GPT-OSS, it’s another bloodbath. Jalapeño’s iso-interactivity throughput per MW is nearly double GB200’s highest throughput point and more than 50x GB200’s concurrency 1 point. The higher concurrency Jalapeño points use EP8.

These results are impressive! However, we have to nitpick: they’re just 8k1k, a much easier workload to tune for, and there are no AgentX runs yet. As mentioned in our AgentX article, multiturn, long context workloads stress much more aspects of the serving stack, such as routers and prefix cache. Many more optimizations are needed to excel in agentic workloads. Read more about this in the AgentX article.

[

](https://newsletter.semianalysis.com/p/agentx-inferencexv3-does-cuda-moat)

[

AgentX - InferenceXv3: Does CUDA Moat Hold up in Agentic Inferencing?

](https://newsletter.semianalysis.com/p/agentx-inferencexv3-does-cuda-moat)

Cam Quilici, Bryan Shan, and 5 others

·

Aug 24

[

Read full story

](https://newsletter.semianalysis.com/p/agentx-inferencexv3-does-cuda-moat)

All these results were gathered on the A0 stepping of Jalapeño, just 9 months into the program. But there is already a B0 stepping that is currently in the fab! B0 has optimizations that deliver roughly a 25% perf-per-watt improvement over the earlier A0 silicon. Specifically, the B0 stepping delivers 13.4 PFLOPs of MXFP4 on a single reticle-sized compute die that is manufactured on TSMC’s N3P. This compares to 17.5 PFLOPs of dense Rubin NVFP4 for a single Rubin compute die that is similar size and on the same node.

This is more respectable considering Jalapeño’s TDP is only 700W compared to Rubin’s at 900-1,150W per compute die. As Jalapeño is geared towards inference rather than training, it is understandable that OpenAI doesn’t need to push TDPs higher to maximize FLOPs, but regardless the above shows that Jalapeño delivers respectable peak theoretical FLOPs.

When compared directly to other accelerators, Jalapeño has the highest HBM bandwidth per watt, and the highest FLOPs per watt, comparable to the 1,800W Rubin Max-Q configuration:

Off-package I/O is provided by an N3E I/O chiplet with 32 lanes of 800G SerDes, for the compute fabric, with 24 lanes (600GB/s) being used for local scale-up within the rack, and 8 lanes (200GB/s) for global scale-up which is the 2,048 XPU multi-rack domain. PCIe Gen 5 is used for system I/O to connect to the x86 host CPU.

Jalapeño will ship with HBM4, making this chip one of the relatively early adopters after Nvidia and AMD, even beating the established TPU and Trainium programs. As one of the key architectural principles behind Jalapeño is getting the most out of HBM bandwidth, settling for anything but the best HBM would run counter to that goal. This results in 15.4TB/s of memory bandwidth per package which bests all the other accelerators shipping that are using HBM3E. The 15.4TB/s bandwidth shows its HBM4 can hit 10Gbps pin speeds, which would give it a slight edge over the 9.6Gbps Nvidia is getting out of its HBM4 in Rubin. The HBM is likely provided by Samsung.

OpenAI taped out Jalapeño in November 2025, or more specifically, this was a tape out of the CoWoS design, not just the top die silicon. Within 9 months of that Nov 2025 tapeout, and with only 3 months of bring-up on actual silicon, OpenAI has already delivered very good results with Jalapeño. This is all the more impressive as the team is starting from zero on the software stack.

Meanwhile, Rubin’s CoWoS tape out was completed in October 2025, a month earlier, and yet the only early results we have seen are from CoreWeave’s engineering samples. Nvidia has not let us test and release benchmarks in the same way that OpenAI has, indicating their chip software is still immature. The CUDA moat is potentially dead given how fast OpenAI can bring up new models on their silicon.

They are still far from optimized and we can see that generally Jalapeño has delivered better numbers. We don’t think that Nvidia hardware is inferior, but more so that Jalapeño’s software bring-up has progressed more quickly than Nvidia’s. This speaks to the power of hardware/software co-design, which is the main area where a cracked frontier lab ASIC team can excel over more established merchant silicon players. Counterintuitively, starting from scratch may also have benefited OpenAI as it could make clean-sheet architectural decisions without worrying about backwards compatibility or older software versions.

While OpenAI has engineering samples of Jalapeño, production is currently scheduled to gradually ramp over 2027 with most of the output currently scheduled for the end of next year. For more details of unit volumes and ASPs, see the SemiAnalysis Accelerator Model.

Suffice to say, OpenAI Jalapeno is a real high volume ASIC.

When compared against Rubin’s timeline, Jalapeño’s is shockingly quick. As shown earlier, Jalapeño’s results beat Rubin’s despite Rubin’s head start.

Digging into the architecture now, the chip’s matrix engine uses MXFP numerical formats and a weight stationary systolic array, similar to TPU. But when compared directly to TPU, it has support for smaller shapes / dimensions, meaning that it doesn’t have weird performance cliffs that get exposed by awkwardly shaped matmuls on bigger systolics.

It also has 64-bit scalar cores and FP32/INT32 vector cores. OpenAI has also invested in redundancy at the tray level and has yield harvesting built in at the core and channel level. They claim that AI assistance in chip design delivered an 8% reduction in SIMD area and a 10% reduction in matrix-engine area during design. While they did not clarify the exact process/voltage/temperature (PVT) conditions, they also mentioned the AI-assisted blocks improved timing and power over the initial blocks.

The Jalapeño architecture design focuses on eliminating fixed latencies and overheads in order to make it possible to get closer to the raw peak flops/bandwidth even for small batches or shapes as compared to other accelerators.

The cores and the HBM are divided into slices, where each core slice has a low-latency local view on its own slice of HBM. Synchronization between slices occurs on a high-bandwidth dedicated collective network. This minimal memory hierarchy already gives Jalapeño a big potential advantage over GPUs, where memory accesses must traverse a complicated memory system, resulting in large latencies that must be amortized or hidden over larger shapes.

This choice is feasible because with careful placement of weights and KVs, synchronization between cores can be restricted to limited, known high-bandwidth comms such as tensor-parallel communication that can be overlapped with compute.

There is also an additional general NoC which is used for general comms and to access the scale-up network. In general OpenAI saves huge power and gets big performance gains with a simplified NOC and memory subsystem vs Nvidia and Google.

At the core level, OpenAI describes an out-of-order (OoO) core with an L1 cache. This is a large divergence from the pattern we have seen in other accelerators, all of which instead use software-managed scratchpad commonly paired with some async DMA support. Again, the argument being made here is that this allows Jalapeño to avoid fixed overheads such as barrier latencies, which on other accelerators (such as GPUs) need to be hidden or amortized over with higher work per core, and make it harder to get close to the raw peak bandwidth/flops.

The tradeoff is that Jalapeño therefore relies on good prefetching to ensure timely arrivals of memory requests, which is less predictable and more difficult to reason about. However, with Codex in a good harness with access to detailed tracing, it is likely that finding the optimal kernel with the best prefetching for a given shape requires little human intervention. We think that is exactly what OpenAI has done to bring up DeepSeek R1, Kimi K2.5, and GPT-OSS so quickly.

The cores also have support for “small” matrix dimensions, which (depending on how small) should make it more general across different model and batch dimensions less sensitive to matrix dimension alignment, padding overhead, and tiling inefficiency. For instance, TPUs, Trainium, and Etched chips have very large systolic arrays which can require large batches or exactly-divisible model dimensions to avoid tiling inefficiencies.

With Jalapeño, OpenAI has focused on eliminating fixed latencies in the system to allow for as-close-to-roofline performance as possible across all areas of the pareto curve. In theory, this could give them advantages over the GPU at multiple operating points:

• Much better upper-bound performance on low-latency/small-batch inference, which on GPUs is limited by many fixed overheads such as launch latencies, barrier latencies, memory system latency

• Some potential to achieve closer to the hardware roofline even for large-batch or long-context

This comes with the caveat that even if the upper-bound performance is available in theory, it may be more difficult to realize that performance for real kernels. So it seems the approach is:

1. Design for the highest upper-bound performance across all workload shapes

2. Let Codex do the tedious work of finding the kernels that achieve that upper bound

Judging by the extremely fast turnaround for the OpenAI team to bring up InferenceX workloads on Jalapeño, we are optimistic about this approach.

If Jalapeño is a success, it will be a strong signal that the industry’s obsession over programming models and perfect, universal compilers are invalidated by frontier AI models.

OpenAI writes Jalapeño kernels like assembly. Each kernel gets hand-tuned code, some running to ~3,000 lines, backed by correctness checks and a custom sanitizer. Early kernel work was human-in-the-loop rather than fully automated, but this shifted with a more scaled-up, internal version of Codex, one which OpenAI plans to pitch to enterprise customers. The internal serving engine is called “Teacup”. Interestingly, OpenAI had no internal implementation of MLA kernels until they benchmarked DeepSeek with InferenceX. The ability for Codex to write functional and efficient kernels so quickly (without any of OpenAI’s kernel engineering team intervening) shows the software pipeline’s developmental ability.

OpenAI programs Jalapeño with Gluon. Gluon is OpenAI’s kernel programming language. Built on top of Triton, Gluon preserves Triton’s SPMD (Single Program Multiple Data) programming model, but it exposes low-level programming abstractions. For example, for NVIDIA GPUs, it offers APIs that map to PTX instructions, including MMA instructions, TMA instructions, mbarrier mechanisms, and many more. The most unique abstraction Gluon provides is the layout. Generally speaking, a layout defines a mapping between a hardware resource (e.g. 5th register of warp 9) and a tensor element (e.g. tensor element on row 6 column 7). Gluon’s layout abstraction is based on Linear Layouts, a type of layout algebra OpenAI invented. Linear Layouts mathematically formalizes what a layout is and provides tools to operate on layouts. This enables many features, such as provably correct layout conversions and optimal memory swizzling.

In terms of Jalapeño’s programming model, each Gluon program maps to a persistent thread. We believe this hints that Jalapeño suits the persistent kernel programming pattern, where each program executes on multiple tiles, and the programmer, rather than the hardware scheduler, assigns the work. OpenAI mentioned TensorInfo, an abstraction that explicitly encodes layouts. This is likely the set of layouts designed for Jalapeño, which will be powered by Linear Layouts. Finally, each core offers data prefetching and decoupled out-of-order units. We believe this indicates a data flow-like programming model, where programmers explicitly manage semaphores. For example, a user might program a wait on a prefetched data, which is locked behind a semaphore.

In a weird twist of fate, OpenAI models like GPT 5.6 Sol, which currently run on NVIDIA GPUs, have been used to design a chip that poses a real threat to the CUDA moat - NVIDIA’s own GPUs are helping usher in their potential successor in real time.

Comparing across time, we can also see Jalapeño’s developmental pace, achieving more than 2x throughput improvements at certain interactivities in less than 2 weeks. Each tarball we get from the Jalapeño team has a world of wonders inside.

Not only did kernel performance improve, in the span of 8 days, the Jalapeño team enabled TP32, building on the previous TP8 configs and expanding beyond a single system to get a full rack-scale config running on a large model. This is a really impressive pace of development.

To validate performance before committing to real hardware runs, OpenAI also has a simulator “chilisim” accurate to within 5% of measured hardware, using a fixed-width trace bus. Tracing on A0 was limited but has improved substantially on B0, likely with inputs from actual runs on A0 silicon. Engineers have demoed the Codex CLI running an internal model, nicknamed “Raiku” or “5.3 Codex Spark”, at 1.2ms TPOT.

The team also showed off Codex-written demos running directly on the chip: Doom at 36 FPS, an FP32 fluid-dynamics simulation, and a “Liquid Light” mouse-drag visualization.

On the model side, OpenAI’s internal megakernel approach, nicknamed “gigakernel”, is built around a single megakernel that loops on-device. The team is also leaning further into test-time compute strategies, with internal interest specifically in how to coherently use 1 million rollouts.

To Disagg or Not to Disagg, ‘tis the Question

We mentioned earlier that OpenAI is not using prefill decode disaggregation on these chips. This came as a surprise to us, as NVIDIA and AMD GPU performance benefits significantly from PDD, even on homogenous hardware. Let’s dig into why the Jalapeño team went this way.

Prefill-decode disaggregation (PDD) looks attractive when the workload is frozen. Prefill and decode stress hardware differently, so assigning each phase to a separately tuned pool can improve efficiency at one chosen input/output ratio. Production traffic, however, does not stay at that ratio. Input and output sequence lengths, concurrency, cache-hit rates, speculative-acceptance rates, and latency targets all move throughout the day.

Once devices are divided into prefill and decode pools, too much prefill demand leaves decode chips idle while requests queue. But too much decode demand does the opposite. The operator must continuously predict the right split, provision spare capacity on both sides, and rebalance a system whose ideal ratio is always moving.

In a unified system, some resources may be underused during a particular phase, but every device remains available to serve the next request. In a disaggregated system, an entire chip can sit idle simply because it belongs to the wrong pool. Local utilization looks better, but global utilization can be bad.

Disaggregation also breaks locality. The prefill worker produces a large KV cache that the decode worker immediately needs, so the system must transfer that state across the network before generation can continue. That adds bandwidth consumption, synchronization, queueing, and another failure domain. The cost also rises with input sequence length because KV cache grows. However, avoiding the movement of KVs is largely a latency optimization; being willing to move some KVs around can allow for increased hardware utilization at the expense of some per-request latency.

A fungible fleet shifts capacity between latency-sensitive requests and throughput-oriented batches, while a fixed split strands hardware whenever the traffic mix changes. Moreover, context length changes the balance between attention and FFN work, making any fixed hardware ratio efficient only near its design point.

The same constraint applies to speculative decoding. A draft model has to feed candidate tokens to the verifier with extremely low latency. Separating the two across specialized pools turns a tightly coupled decoding loop into a distributed protocol. The extra communication and coordination can consume the latency saved by drafting. Keeping both models on the same devices and low-latency fabric preserves the locality that makes speculation worthwhile in the first place.

However, disaggregation can still win where demand is sufficiently large, stable, and predictable, particularly when conventional GPUs need large phase-specific batches to reach good throughput. But it is not free lunch.

The Jalapeño System at the rack unit level consists of a CPU host rack and an ASIC rack. The host rack houses 16 host CPU trays named “Katsu,” each corresponding to one of the 16 ASIC trays, named “Vindaloo,” to the right of the Katsu. Each host houses two Turin-class AMD EPYC CPUs with 1.5TB of DRAM, 2x E1.S, and 2x M.2 SSDs per rack. Each tray is also specced with 400G (2x200G) frontend networking. Each Katsu tray connects to each Vindaloo tray via 8 external PCIe DAC cables that run horizontally across the rack at the front. The system level design is done in partnership with Celestica.

The ASIC rack consists of 16 Vindaloo trays and 8 scale up switch trays (6 for local + 2 for global), named “Chana.” Each Vindaloo tray consists of 8 Jalapeño ASICs, making up a total of 128 Jalapeño ASICs per rack. The ASICs are connected to each of the Chana switch trays via a copper cable backplane, just like that of Nvidia’s Oberon. The scale up topology is split into a local domain of 128 ASICs within the rack and a global domain connecting up to 16 racks or 2,048 ASICs. We will explain the bandwidth and the topology in more detail below.

Power provisioning to a sidecar host rack draws roughly 50kW provisioned (31kW in production), and the ASIC rack draws 130kW, making the total two rack system roughly 160kW. That’s basically a double-wide GB300 rack in terms of power draw.

OpenAI can connect up to 2,048 Jalapeño XPUs within a single scale-up network. The scale-up network consists of two domains, a local domain connecting all 128 XPUs over backplane within the rack, as well as a global domain connecting 2,048 XPUs over 16 racks using a hybrid of copper and optical interconnect. Each rack consists of 8 Chana switch trays. Six Chana switches in the middle are for the local domain, which come with one 102.4T Tomahawk 6 switch ASIC each. Two Chana switches at the top and bottom of the local switches are for the global domain, which we think could consist of 2x 102.4T Tomahawk 6 switches making up to 204.8T per switch tray.

In the local domain, each of the 128 Jalapeño chips has a per XPU uni-directional bandwidth of 4.8Tb/s and is connected on an all-to-all basis to 6x 102.4Tb/s Tomahawk 6 ASICs. This would amount to 48-differential pair (DP) male and female connector pairs per XPU translating to a total of 6,144DPs worth of passive copper cables per rack used for local scale-up.

For the global domain, 16 racks totaling 2,048 XPUs are connected together via a combination of copper backplane, electrical 204.8T TH6 switch, 1.6T transceivers, and optical circuit switch. Each XPU has a uni-directional bandwidth of 1.6Tb/s for the global link, which is 16-differential pair (DP) male and female connector pairs per XPU for the backplane between the XPU and the global switch. Bandwidth exiting each global switch tray of 2 ASICs each is split between the backplane and front panel optics.

Between local domain and global domain, backplane connector count per rack comes up to 64 DPs per XPU and a total of 8,192 DPs worth of passive copper cables per rack.

The global domain adopts a rail-only architecture consisting of 8-rails across the global domain. We think OpenAI routes optical links in the global domain via Optical Circuit Switches (OCS) installed in every rack. For every XPU, 1.6Tb/s of global bandwidth will travel to the global switch tray over the copper backplane. This then exits the switch through the front panel via 1.6T transceivers, which go to the passive optical switch before exiting the rack. This expands the scale-up world size to 2,048 XPUs combining 16 racks of 128 XPUs each.

Because scale-up networking is only about 10% of total system cost, that flexibility buys valuable optionality for future 10–20 trillion parameter models or 2–4 million token context windows. On deployment, OpenAI is partnering with neoclouds and is gathering reliability data with datacenter partners through January while optimizing dock-to-rack rollout time.

Next, we talk about the future of Jalapeño, whose first production token is coming soon. The next goal is 100MW, and the hurdles will mostly be hardware: How much can they produce, how well can they deploy and operate datacenters, how do they handle monitoring, and resiliency, etc. The software is already proven, and with internal models, every software headstart is easily caught up to. Behind the paywall we will discuss implications for NVIDIA, AMD, Cerebras, and other chip companies who have signed deals with OpenAI in the coming years.

Direct Implications for NVIDIA, AMD and Cerebras

Prof Agus Kartono dari IPB University kembangkan pemodelan biofisik untuk diagnosis dini dan terapi presisi penyakit Parkinson serta Alzheimer

Penunjukan mantan Menteri Ekonomi Kreatif itu, sebagai Presiden Komisaris MUTU, disahkan dalam Rapat Umum Pemegang Saham Luar Biasa Senin (24/8/2026)

Manajer Persija, Ardhi Tjahjoko, berharap berbagai hal positif yang telah terbentuk selama pemusatan latihan tidak berhenti setelah tim kembali ke Jakarta.

A research duo, once pitched to lead the Jeff Bezos-backed Project Prometheus, unveiled what they've built on their own: an AI model that can take ​in and produce information so vast that a lone computer might not handle it.

Workers are "time traveling" on LinkedIn to add AI skills to roles they left years ago, likely to get a leg up in a tough job market, researchers say.

Indonesia jadi salah satu target normalisasi Israel.

Komplotan maling membobol minimarket di Deli Serdang, mencuri barang senilai Rp 29 juta. Polisi tangkap satu pelaku, sementara tiga lainnya masih diburu.

Joe Mazzulla berhasil membawa Boston Celtics menjadi juara NBA 2024. Tetapi Mazzulla enggan menyandang status pelatih juara. Bagi Mazzulla, satu cincin juara belum cukup untuk menentukan nilai seorang pelatih.

NVIDIA NemoClaw's Windows-host Ollama path exposes an unauthenticated API that DNS rebinding could use to plant persistent model instructions.

5 Motor Touring Bekas yang Bandel dan Murah, Cocok untuk Pemula

Guy Ritchie's mob drama The Gentlemen has scored a Season 3 renewal — and dropped an explosive first trailer for Season 2, which hits Netflix in September.

WhatsApp adds multiple passkeys per account, full-password two-step verification, and richer Android caller context for unknown contacts.

Ancaman AS untuk menjatuhkan sanksi terhadap entitas China terkait perdagangan dengan Iran kecil kemungkinannya akan terwujud karena adanya kombinasi hambatan diplomatik...

Mataram (Suara NTB) – Partai Demokrat lebih memilih fokus melakukan konsolidasi internal untuk memperkokoh mesin politik partai untuk persiapan menghadapi pemilu 2029. Dengan demikian, seperti apapun skenario dinamika politik yang akan dihadapi, seiring dengan pembahasan undang-undang pemilu, Demokrat dipastikan siap menghadapinya. Hal itu disampaikan oleh Sekretaris I Badan Pembinaan Organisasi, Keanggotaan, dan Kaderisasi (BPOKK) DPP […]

Kecelakaan di Tol Tangerang-Merak, truk jatuh ke jalur kereta api. Foto udara kondisi truk yang mengalami kecelakaan di Unyur, Kota Serang, Banten, Selasa (25/8/2026). Truk tersebut ...

Menanggapi sanksi terbaru Amerika Serikat terhadap Iran, Kementerian Luar Negeri China menyatakan kerja samanya dengan Teheran senantiasa mematuhi kerangka hukum...

Pada 1908 Grete menjadi salah satu perempuan pertama yang menembus lingkungan akademik yang didominasi laki-laki

Japan automaker falters in longtime stronghold of Southeast Asia amid rise of Chinese EVs

Hari libur nasional pada hari ini, Selasa, 25 Agustus 2026 merupakan hari libur nasional terakhir sampai perayaan Natal mendatang.

See what it takes to convince an AI writing detector that AI-generated text was written by a human.

Enam bulan perang AS dan Iran, lalu lintas kapal komersil di Selat Hormuz masih terbatas sementara ribuan pelaut telantar.

Mahasiswi Universitas Ichsan Gorontalo yang meninggal di Kos-kosan di Kelurahan Dulalowo Timur diduga punya Riwayat penyakit Jantung.

Deteksi dini Parkinson, demensia, dan gangguan saraf dinilai penting seiring bertambahnya usia harapan hidup dan kebutuhan layanan neurologi

Film Mutiny mengisahkan tentang Cole Reed diperankan oleh Jason Statham seorang pria yang terjebak dalam situasi mematikan setelah menyaksikan pembunuhan bosnya